| \documentclass[10pt,twocolumn]{article} |
| \IfFileExists{lmodern.sty}{\usepackage{lmodern}}{} |
| \usepackage{amssymb,amsmath} |
| \usepackage{ifxetex,ifluatex} |
| \IfFileExists{fixltx2e.sty}{\usepackage{fixltx2e}}{} |
| \ifnum0\ifxetex1\fi\ifluatex1\fi=0 |
| \usepackage[T1]{fontenc} |
| \usepackage[utf8]{inputenc} |
| \else |
| \ifxetex |
| \usepackage{mathspec} |
| \usepackage{xltxtra,xunicode} |
| \else |
| \usepackage{fontspec} |
| \fi |
| \defaultfontfeatures{Mapping=tex-text,Scale=MatchLowercase} |
| \newcommand{\euro}{EUR} |
| \fi |
| |
| \IfFileExists{upquote.sty}{\usepackage{upquote}}{} |
| |
| \IfFileExists{microtype.sty}{\usepackage{microtype}}{} |
| \IfFileExists{geometry.sty}{\usepackage[letterpaper,top=0.8in,bottom=0.9in,left=0.75in,right=0.75in]{geometry}}{} |
| \usepackage{array,tabularx,booktabs} |
| \usepackage{graphicx} |
| \ifxetex |
| \usepackage[setpagesize=false, |
| unicode=false, |
| xetex]{hyperref} |
| \else |
| \usepackage[unicode=true]{hyperref} |
| \fi |
| \hypersetup{breaklinks=true, |
| bookmarks=true, |
| pdfauthor={Tyler Baumgartner, Brandon Tai, Lisa Kaelin-Martin, Luc Debaupte, Bill Wang, Yi Zhong}, |
| pdftitle={VoiceCodeBench: Evaluating Exact Structured-Token Recovery in Automatic Speech Recognition}, |
| colorlinks=true, |
| citecolor=blue, |
| urlcolor=blue, |
| linkcolor=magenta, |
| pdfborder={0 0 0}} |
| \urlstyle{same} |
| \setlength{\parindent}{10pt} |
| \setlength{\parskip}{0pt} |
| \setlength{\emergencystretch}{3em} |
| \setcounter{secnumdepth}{0} |
| \setlength{\tabcolsep}{3pt} |
| \setlength{\columnsep}{0.25in} |
| \renewcommand{\arraystretch}{1.08} |
| \newcolumntype{L}{>{\raggedright\arraybackslash}X} |
|
|
| \title{VoiceCodeBench: Evaluating Exact Structured-Token Recovery in Automatic Speech Recognition} |
| \author{Tyler Baumgartner, Brandon Tai, Lisa Kaelin-Martin, Candice Fan,\\ |
| Luc Debaupte, Bill Wang, and Yi Zhong\\ |
| Besimple AI, San Mateo, CA\\ |
| \texttt{\{tyler,yi\}@besimple.ai}} |
| \date{} |
|
|
|
|
| \begin{document} |
|
|
| \maketitle |
|
|
| \begin{abstract} |
|
|
| Automatic speech recognition (ASR) systems are commonly evaluated with word |
| error rate (WER), yet many voice workflows depend on exact written |
| values for key entities. A transcript can appear fluent and achieve low WER |
| while corrupting an identifier, path, or measured quantity that a |
| downstream system must parse or store. When building voice agents, these |
| errors are not merely transcription defects. They can become wrong tool |
| arguments, invalid database fields, misrouted requests, or unsafe commands |
| while the transcript still looks plausible. |
|
|
| We introduce \textbf{VoiceCodeBench}, a benchmark for evaluating exact |
| structured-token recovery in English ASR\@. VoiceCodeBench contains 300 |
| human-recorded workplace segments across eight workflow domains and |
| 1,482 audited target entities across 26 entity types, each with a |
| canonical written form recoverable from the spoken evidence. |
|
|
| VoiceCodeBench uses a raw-audio-only evaluation protocol. Systems |
| receive only audio bytes, with no additional context or metadata. Alongside |
| WER, we introduce entity-sensitive measures, including Canonical |
| Token/Entity Match (CTEM), Task Success Rate (TSR), and per-type exact |
| recovery, to compare broad transcript accuracy with exact recovery of |
| workflow-critical values. |
|
|
| Across 12 baseline ASR systems, lower WER generally corresponded to better |
| structured-token recovery, but it did not fully determine it: Spearman |
| correlation was -0.73 between WER and CTEM and -0.73 between WER and TSR\@. |
| Even the strongest baseline ASR system by TSR reached only 68.7\% TSR, |
| meaning that nearly one third of recordings still contained at least |
| one unrecovered workflow-critical value. These results show that WER is |
| an informative transcript-quality diagnostic, but entity-sensitive |
| metrics are needed to measure whether ASR output preserves the |
| identifiers, commands, paths, measurements, and other exact values that |
| production systems must parse, route, store, compare, or execute. |
| \end{abstract} |
|
|
| \section{Introduction}\label{introduction} |
|
|
| Speech interfaces increasingly mediate workflows that require exact |
| written values, such as contact details, file paths, account numbers, |
| and measured quantities. In such settings, a transcript is not only a |
| readable approximation of speech. It is software input that may be |
| stored, routed, compared, or executed. A sentence can look fluent while |
| still being unusable for the application that consumes it. |
|
|
| This creates a mismatch between common automatic speech recognition |
| (ASR) evaluation and production requirements. Word error rate (WER) |
| remains the standard summary metric for broad transcription quality |
| (National Institute of Standards and Technology, 2021), but it treats |
| word-level edits largely as interchangeable. A missing article can |
| count much like a corrupted structured value, even though only the latter |
| may misroute a support request, populate a record with the wrong |
| identifier, apply the wrong amount or unit, or execute the wrong tool |
| call. For structured tokens, punctuation, casing, separators, |
| token boundaries, digits, and spoken-to-written normalization can be part |
| of the value itself. A low-WER transcript can therefore be readable to a |
| person and still unsafe for the software system that consumes it. |
|
|
| Existing ASR benchmarks make broad progress measurable across many |
| sources of difficulty. LibriSpeech standardizes read audiobook speech, |
| Common Voice broadens speaker/language coverage through crowdsourcing, |
| GigaSpeech scales English ASR across web domains, FLEURS emphasizes |
| multilingual transfer, and AMI captures multi-party meeting speech |
| (Panayotov et al., 2015; Ardila et al., 2020; Chen et al., 2021; |
| Conneau et al., 2022; Carletta, 2007). These resources support |
| measurement of acoustic robustness, domain and language generalization, |
| and conversational transcription. A parallel line of work connects ASR |
| quality to meaning and application utility, including |
| spoken-language understanding accuracy, semantic distance, downstream |
| usability, named-entity recognition over ASR transcripts, and contextual |
| ASR (Wang et al., 2003; Kim et al., 2021; Roy, 2021; Szymanski et al., |
| 2023; Wang et al., 2025). However, these evaluations usually make |
| transcript fidelity, semantic similarity, named-entity recognition, or |
| context use the primary target. They do not directly answer a narrower |
| production question: given raw audio alone, did the ASR output preserve |
| enough evidence for every workflow-critical written value to be recovered |
| exactly? |
|
|
| VoiceCodeBench extends this foundation by making exact written-value |
| recoverability the unit of analysis. The benchmark focuses on ordinary |
| workplace domains where structured values are often the purpose of the |
| voice interaction rather than incidental edge cases. It is constructed |
| entity-first, with target entities allocated before transcripts are |
| written, producing controlled coverage by type, domain, and difficulty. |
| Each entity has a canonical written form recoverable from the acoustic |
| evidence, so evaluation can ask whether the transcript preserves the |
| specific value that an application would need to store, route, compare, |
| or execute. |
|
|
| VoiceCodeBench contains 300 human-recorded English workplace segments |
| with 1,482 target entities across 26 entity types and eight domains. |
| Under its raw-audio-only protocol, 12 ASR systems receive audio bytes |
| without target entities, domain labels, candidate values, prompts, custom |
| vocabulary, or other metadata. WER is reported alongside Canonical |
| Token/Entity Match (CTEM), Task Success Rate (TSR), and per-type exact |
| recovery. These metrics make different deployment questions visible. |
| WER measures broad transcript quality, CTEM measures value-level recovery |
| load, TSR measures whether a full segment can pass through an automated |
| workflow without repair, and per-type recovery identifies which classes |
| need safeguards such as confirmation prompts, constrained decoding, or |
| post-ASR validation. The benchmark therefore tests not only whether ASR |
| output reads well, but whether it preserves the exact values on which |
| workflow correctness depends. |
|
|
| \section{Methods}\label{methods} |
|
|
| \subsection{Benchmark Design}\label{benchmark-design} |
|
|
| VoiceCodeBench is designed as a test-only benchmark for evaluating |
| exact structured-token recovery in automatic speech recognition (ASR). |
| Each benchmark item contains a human-recorded English audio segment, a |
| reference transcript, item metadata including speaker and audio-quality |
| fields, and a set of target entities whose canonical written forms are |
| recoverable from their spoken form. The benchmark focuses on compact |
| workplace-style utterances containing values that downstream |
| applications may parse, route, store, compare, or execute. |
|
|
| During evaluation, ASR systems receive only the audio file |
| at inference time. This scope reflects a common pattern where |
| developers submit audio to an ASR provider and use the returned |
| transcript directly or with lightweight downstream processing. |
|
|
| The dataset contains 300 human-recorded English segments |
| across eight workflow domains, summarized in |
| Table~\ref{tab:workflow-domains}. The domain distribution is not |
| intended as an estimate of how often these workflows occur in |
| production speech. It is designed to provide enough examples of each |
| structured-token failure mode for per-type analysis. |
|
|
| \begin{table*}[t]\centering\footnotesize |
| \caption{Workflow-domain distribution.}\label{tab:workflow-domains} |
| \begin{tabularx}{\textwidth}{lrL} |
| \toprule |
| Domain & Recordings & Purpose |
| \\ |
| \midrule |
| Contact/routing & 45 & Callback numbers, extensions, emails, postal |
| addresses, spelled names, routing teams |
| \\ |
| Technical/IT/dev & 55 & Commands, flags, files, URLs, IPs, ports, |
| versions, symbols, environment variables |
| \\ |
| Retail/logistics/order & 45 & SKUs, serials, tracking IDs, quantities, |
| addresses, returns, subscriptions |
| \\ |
| Finance/billing & 40 & Invoices, account numbers, currency, percentages, |
| rates, dates, payment details |
| \\ |
| Healthcare/admin & 35 & Appointments, measurements, dosages, record |
| numbers, referrals, insurance terms |
| \\ |
| Legal/insurance/government & 35 & Claims, policies, case IDs, exhibit |
| labels, formal references, addresses |
| \\ |
| Education/workplace & 25 & HR, courses, rooms, employee IDs, internal |
| helpdesk, facilities |
| \\ |
| Dense mixed stress & 20 & High entity load and cross-type interference |
| under compact spoken conditions |
| \\ |
| \bottomrule |
| \end{tabularx}\end{table*} |
|
|
| The benchmark includes 1,482 target entities across 26 entity types, |
| with an average of 4.94 target entities per recording. Gold acoustic |
| transcript lengths range from 95 to 206 words depending on difficulty |
| band. Each recording is assigned a difficulty band that combines entity load, |
| entity complexity, transcript length, and expected recovery challenge. |
| Table~\ref{tab:difficulty-bands} summarizes the band design. |
|
|
| \begin{table}[t]\centering\footnotesize |
| \caption{Difficulty-band design.}\label{tab:difficulty-bands} |
| \begin{tabular}{lrll} |
| \toprule |
| Difficulty & Recordings & Entity load & Acoustic transcript length |
| \\ |
| \midrule |
| Light & 30 & 3 entities & 95--130 words |
| \\ |
| Standard & 114 & 4 entities & 103--169 words |
| \\ |
| Dense & 93 & 5--6 entities & 122--177 words |
| \\ |
| Stress & 63 & 7--8 entities & 156--206 words |
| \\ |
| \bottomrule |
| \end{tabular}\end{table} |
|
|
| All scenarios and structured values are synthetic. The |
| \hyperref[ethical-and-privacy-considerations]{Ethical and Privacy |
| Considerations section} describes the privacy constraints used when |
| constructing these values. |
|
|
| VoiceCodeBench organizes target entities by recovery behavior rather |
| than by workflow domain alone. The taxonomy contains 26 entity types |
| grouped into six broad categories (Table~\ref{tab:entity-taxonomy}). |
|
|
| \begin{table*}[t]\centering\footnotesize |
| \caption{Entity taxonomy used for structured-token scoring.}\label{tab:entity-taxonomy} |
| \begin{tabularx}{\textwidth}{lL} |
| \toprule |
| Group & Entity types |
| \\ |
| \midrule |
| Contact and routing & \texttt{email\_address}, \texttt{phone\_number}, |
| \texttt{phone\_extension}, \texttt{person\_or\_team\_name}, |
| \texttt{postal\_address} |
| \\ |
| Network and web & \texttt{url}, \texttt{ip\_address}, |
| \texttt{port\_number} |
| \\ |
| Code and system & \texttt{command}, \texttt{cli\_flag}, |
| \texttt{file\_path}, \texttt{environment\_variable}, |
| \texttt{code\_symbol}, \texttt{version} |
| \\ |
| Identifiers & \texttt{reference\_id}, \texttt{product\_code}, |
| \texttt{account\_or\_record\_number} |
| \\ |
| Numeric and measurement & \texttt{currency\_amount}, |
| \texttt{percentage}, \texttt{measurement}, \texttt{plain\_number}, |
| \texttt{date}, \texttt{time} |
| \\ |
| Language form & \texttt{acronym\_or\_initialism}, |
| \texttt{spelled\_sequence}, \texttt{domain\_term} |
| \\ |
| \bottomrule |
| \end{tabularx}\end{table*} |
|
|
| The taxonomy captures values that are both exactness-sensitive and critical to |
| real-world ASR-backed production workflows. Some entities, such as phone |
| numbers or dates, may be recoverable under conventional formatting |
| normalization. Others, such as file paths or command-line flags, treat |
| punctuation and separators as part of the value itself. |
|
|
| VoiceCodeBench is not intended as a general-purpose ASR corpus or a |
| training set. It is released as a diagnostic evaluation resource. |
|
|
| \subsection{Entity and Transcript |
| Generation}\label{entity-transcript-generation} |
|
|
| VoiceCodeBench was constructed entity-first. For each item, we first |
| specified the workflow domain, difficulty band, target entity count, |
| and target entity types. We then generated unique synthetic entities for |
| the required slots, assigned each value both an acoustic form and a |
| canonical form. Finally, we wrote a workplace-style transcript around |
| that planned entity bundle. This ordering prevents the dataset from |
| merely extracting whichever structured values happen to appear in |
| free-form scripts, and it allows coverage to be controlled across |
| entity types, workflow domains, and difficulty bands. |
|
|
| Generation was LLM-assisted but constraint-driven. We used a |
| repository-aware LLM workflow to draft and check items under fixed |
| metadata, entity, and validation constraints. Implementation details are |
| provided in \hyperref[appendix-generation-tooling]{the generation-tooling |
| appendix}. This workflow helped maintain the benchmark |
| metadata structure, populate the template, acoustic, and canonical |
| transcript layers, and enforce the requested domain, difficulty, and |
| entity-type constraints. Candidate items were accepted only after |
| validation and review for domain fit, naturalness, uniqueness, entity |
| consistency, and recoverability. |
|
|
| The acoustic form represents what the speaker is expected to say. It |
| may include spoken symbols, spelling cues, casing instructions, or |
| formatting instructions. The canonical form represents the exact written |
| value that a downstream application must recover. This distinction makes |
| explicit the gap between what is said and what software needs to |
| consume. |
|
|
| \[ |
| \begin{array}{@{}rcl@{}} |
| \text{\footnotesize ``double dash dry dash run''} & \rightarrow & |
| \text{\footnotesize\texttt{-\kern0pt-dry-run}} \\ |
| \text{\footnotesize ``all caps database underscore URL''} & \rightarrow & |
| \text{\footnotesize\texttt{DATABASE\_URL}} |
| \end{array} |
| \] |
|
|
| The acoustic form alone must provide enough information for a careful |
| listener to infer the intended canonical value. An entity is rejected or |
| revised if its canonical value is not uniquely recoverable from the |
| intended acoustic form. For example, a target requiring |
| \texttt{DATABASE\_URL} must include sufficient spoken evidence for the |
| underscore and casing convention. |
|
|
| The scenarios and structured values are synthetic, while the released |
| audio is human-recorded. Synthetic content avoids exposing real contact |
| details, accounts, credentials, or operational systems. Email addresses |
| and URLs use reserved documentation domains and controlled domains. |
| Phone numbers use fictional NANP \texttt{555\mbox{-}0100} through |
| \texttt{555\mbox{-}0199} numbers with varied area codes. Public-looking IPv4 |
| addresses use documentation ranges, while internal network examples use |
| private address ranges. Postal addresses, account numbers, product |
| codes, reference IDs, and workflow scenarios are synthetic. |
|
|
| The dataset may include public organization, product, or platform names |
| when they function as ordinary workplace vocabulary, but these names are |
| not paired with real private contact records or live routable |
| information. This allows the benchmark to test recognition of common |
| workplace terms while keeping sensitive fields synthetic or reserved. |
|
|
| \subsection{Recording and Verification}\label{recording-verification} |
|
|
| Each transcript is recorded by a human speaker reading one compact |
| workplace-style segment. Speakers read the acoustic transcript layer, |
| which renders target entities in the form intended to be spoken rather |
| than as their canonical scoring values. Speakers are instructed to read |
| naturally but clearly, preserving dictated punctuation phrases, spelling |
| sequences, casing cues, and formatting instructions. The intended style is |
| deliberate workplace dictation rather than theatrical performance or |
| casual conversation. |
|
|
| Recordings are collected remotely through a crowdsourcing platform from |
| paid contributors who consent to dataset use and release. After recording, |
| each audio file is human audited. Files with severe audio quality |
| issues or that contain spoken errors are rejected and re-recorded. |
| Accepted audio files are released without additional post-processing. |
|
|
| \subsection{Transcription Protocol}\label{transcription-protocol} |
|
|
| All ASR systems receive only the raw audio file for each |
| recording. Batch systems transcribe each complete file; streaming systems |
| receive chronological audio chunks under a fixed chunking policy. Only final |
| transcripts are scored. Provider defaults are used except for settings |
| needed to select the intended model or English language mode. Provider |
| punctuation, capitalization, and formatting are allowed, but benchmark- |
| specific prompting, custom vocabulary, entity hints, grammar constraints, |
| and post-ASR correction are excluded from the main raw-audio-only evaluation. |
| For each baseline ASR system, we record the provider, model, endpoint or API, batch or |
| streaming mode, evaluation date, and reproduction-relevant inference settings. |
|
|
| \subsection{Entity Extraction}\label{entity-extraction} |
|
|
| Entity extraction asks whether each gold canonical value is recoverable |
| from the ASR transcript. Formatting variation is accepted when it preserves |
| the same value, such as a phone number rendered as an uninterrupted digit |
| sequence or a currency amount rendered as an unambiguous spoken amount. |
| Separators, casing cues, units, digits, token boundaries, and punctuation are |
| treated as value-bearing when they determine the canonical string, as in email |
| addresses, file paths, URLs, environment variables, code symbols, and command-line |
| flags. |
|
|
| The entity extraction policy is applied with an LLM-powered recoverability |
| verifier. The verifier receives the ASR transcript and the target entities, |
| including type, acoustic form, and canonical form, and returns strict JSON |
| with one result per target index, including type, canonical value, present/absent |
| decision, evidence span, and reason. It is instructed to mark an entity |
| present only when the transcript contains enough evidence to recover the exact |
| canonical value, and to reject wrong, missing, extra, or substituted letters, |
| digits, separators, units, dates, times, amounts, or words that change the |
| value. The tracked verifier is the versioned artifact |
| \texttt{openai\_gpt\_5\_5\_v1}, using \texttt{GPT-5.5}. The full verifier |
| system prompt is provided in |
| \hyperref[appendix-verifier-prompt]{the verifier-prompt appendix}. Verifier |
| outputs include evidence and reasons, and edge cases |
| identified during baseline analysis were manually reviewed. As a reliability |
| check, a human auditor reviewed a stratified sample of 200 entity decisions |
| across ASR models and entity types. Agreement with the verifier decisions was |
| 100\%. The reviewed rows are released as |
| \texttt{audit/verifier\_audit\_samples.csv}. |
|
|
| \subsection{Metrics and Scoring}\label{metrics}\label{statistical-reporting}\label{baseline-systems} |
|
|
| VoiceCodeBench includes WER as a broad transcript-accuracy comparison |
| and reports entity-sensitive metrics for workflow-critical value |
| recovery. WER compares the model transcript with the gold acoustic |
| transcript after both strings are lowercased and reduced to word tokens. |
| Punctuation and separators are not scored as standalone WER tokens. |
|
|
| The primary entity metric is Canonical Token/Entity Match: |
|
|
| \[ |
| \mathrm{CTEM} = |
| \frac{N_{\mathrm{correct\ entities}}}{N_{\mathrm{target\ entities}}} |
| \] |
|
|
| CTEM is measured overall and by entity type, workflow domain, difficulty |
| band, and evaluation mode. Task Success Rate measures whether all target |
| entities in a recording are recovered correctly: |
|
|
| \[ |
| \mathrm{TSR} = |
| \frac{N_{\mathrm{successful\ recordings}}}{N_{\mathrm{recordings}}} |
| \] |
|
|
| TSR is intentionally strict. Many workflow segments are not safe to use |
| in production settings if even one required value is corrupted. Overall |
| confidence intervals are calculated using Wilson 95\% binomial intervals. |
| Slice tables are reported as descriptive baseline summaries without |
| confidence intervals. |
|
|
| VoiceCodeBench evaluates both commercial and open ASR systems in batch |
| and streaming settings. The baseline suite covers ASR systems from |
| major commercial providers and open models under the raw-audio-only |
| protocol described above. Table~\ref{tab:baseline-systems} lists the |
| baseline ASR systems used in the benchmark. |
|
|
| \begin{table*}[t]\centering\footnotesize\begin{tabular}{llc} |
| \toprule |
| Provider & Model & Mode |
| \\ |
| \midrule |
| Deepgram & Nova-3 & Batch |
| \\ |
| OpenAI & GPT-4o Transcribe & Batch |
| \\ |
| AssemblyAI & Universal-3 Pro & Batch |
| \\ |
| Google Cloud & Chirp 3 & Batch |
| \\ |
| ElevenLabs & Scribe v2 & Batch |
| \\ |
| Groq & Whisper Large v3 & Batch |
| \\ |
| Amazon Transcribe & Amazon Transcribe Streaming & Streaming |
| \\ |
| Deepgram & Nova-3 & Streaming |
| \\ |
| OpenAI & GPT Realtime Whisper & Streaming |
| \\ |
| AssemblyAI & Universal-3 Pro & Streaming |
| \\ |
| Google Cloud & Chirp 3 & Streaming |
| \\ |
| ElevenLabs & Scribe v2 Realtime & Streaming |
| \\ |
| \bottomrule |
| \end{tabular} |
| \caption{Baseline ASR systems evaluated by VoiceCodeBench.}\label{tab:baseline-systems} |
| \end{table*} |
|
|
| \section{Results}\label{results} |
|
|
| \subsection{Dataset Composition}\label{dataset-composition} |
|
|
| The completed baseline release evaluates the 300-recording benchmark |
| described in Methods. The evaluated audio totals 5.58 hours from 85 |
| unique speakers, and the baseline suite contains 12 baseline ASR systems. |
|
|
| The target-entity distribution is intentionally broad. Entity types such |
| as reference IDs, product codes, dates, acronyms, and spelled sequences |
| appear frequently, while lower-frequency but operationally important |
| classes such as file paths, environment variables, port numbers, |
| percentages, and measurements are explicitly represented. |
| Table~\ref{tab:entity-counts} reports the target-entity counts by type. |
|
|
| \begin{table*}[t]\centering\footnotesize |
| \caption{Target-entity counts by entity type.}\label{tab:entity-counts} |
| \begin{tabularx}{\textwidth}{LrLrLr} |
| \toprule |
| Entity type & Count & Entity type & Count & Entity type & Count |
| \\ |
| \midrule |
| \texttt{reference\_id} & 150 & \texttt{url} & 62 & \texttt{code\_symbol} & 35 |
| \\ |
| \texttt{spelled\_sequence} & 97 & \texttt{measurement} & 61 & \texttt{environment\_variable} & 35 |
| \\ |
| \texttt{product\_code} & 90 & \texttt{phone\_number} & 60 & \texttt{phone\_extension} & 30 |
| \\ |
| \texttt{date} & 89 & \texttt{time} & 60 & \texttt{port\_number} & 30 |
| \\ |
| \texttt{acronym\_or\_initialism} & 85 & \texttt{file\_path} & 51 & \texttt{version} & 30 |
| \\ |
| \texttt{currency\_amount} & 75 & \texttt{command} & 50 & \texttt{ip\_address} & 25 |
| \\ |
| \texttt{account\_or\_record\_number} & 65 & \texttt{percentage} & 50 & \texttt{domain\_term} & 20 |
| \\ |
| \texttt{plain\_number} & 65 & \texttt{cli\_flag} & 44 & \texttt{person\_or\_team\_name} & 18 |
| \\ |
| \texttt{email\_address} & 65 & \texttt{postal\_address} & 40 & & |
| \\ |
| \bottomrule |
| \end{tabularx}\end{table*} |
|
|
| \subsection{Overall ASR and Structured-Token |
| Performance}\label{overall-asr-and-structured-token-performance} |
|
|
| Table~\ref{tab:overall-results} reports overall performance for each |
| baseline ASR system. All scores are percentages. |
|
|
| \begin{table*}[t] |
| \centering |
| \footnotesize |
| \setlength{\tabcolsep}{4.5pt} |
| \begin{tabular*}{\textwidth}{@{\extracolsep{\fill}}l l r r c r c@{}} |
| \toprule |
| Model & Mode & WER $\downarrow$ & CTEM $\uparrow$ & CTEM 95\% CI & TSR $\uparrow$ & TSR 95\% CI |
| \\ |
| \midrule |
| ElevenLabs Scribe v2 & Batch & 16.1 & \textbf{91.6} & 90.0--92.9 & 67.7 & 62.2--72.7 |
| \\ |
| Deepgram Nova-3 & Batch & \textbf{8.6} & 90.9 & 89.3--92.3 & \textbf{68.7} & 63.2--73.7 |
| \\ |
| Deepgram Nova-3 & Streaming & 9.6 & 88.9 & 87.2--90.4 & 61.7 & 56.1--67.0 |
| \\ |
| Google Cloud Chirp 3 & Batch & 25.6 & 88.8 & 87.1--90.3 & 60.3 & 54.7--65.7 |
| \\ |
| OpenAI gpt-realtime-whisper & Streaming & 16.3 & 88.7 & 87.0--90.2 & 59.3 & 53.7--64.7 |
| \\ |
| Whisper large-v3 & Batch & 23.1 & 87.6 & 85.8--89.2 & 54.3 & 48.7--59.9 |
| \\ |
| OpenAI gpt-4o-transcribe & Batch & 24.6 & 87.4 & 85.6--89.0 & 55.7 & 50.0--61.2 |
| \\ |
| Google Cloud Chirp 3 & Streaming & 25.5 & 86.1 & 84.2--87.8 & 50.3 & 44.7--56.0 |
| \\ |
| AssemblyAI Universal-3 Pro & Batch & 25.2 & 84.5 & 82.6--86.3 & 50.3 & 44.7--56.0 |
| \\ |
| ElevenLabs Scribe v2 realtime & Streaming & 24.9 & 84.0 & 82.1--85.8 & 46.3 & 40.8--52.0 |
| \\ |
| AssemblyAI Universal-3 Pro & Streaming & 25.2 & 78.3 & 76.1--80.3 & 33.0 & 27.9--38.5 |
| \\ |
| Amazon Transcribe & Streaming & 25.6 & 75.2 & 73.0--77.4 & 33.7 & 28.6--39.2 |
| \\ |
| \bottomrule |
| \end{tabular*} |
| \caption{Overall baseline performance on 300 recordings and 1,482 target entities.}\label{tab:overall-results} |
| \end{table*} |
|
|
| Across baseline ASR systems, WER ranged from 8.6 to 25.6, while CTEM |
| ranged from 75.2 to 91.6 and TSR ranged from 33.0 to 68.7. The main |
| empirical result is that transcript-level accuracy and structured-token |
| recovery are correlated, but the association is incomplete. |
|
|
| To test whether ordinary transcript accuracy predicts structured-token |
| correctness, we compare WER against CTEM and TSR across systems |
| (Figure~\ref{fig:wer-entity-scatter}). WER has Spearman correlation |
| $\rho=-0.73$ with CTEM and $\rho=-0.73$ with TSR\@. The negative signs |
| reflect the opposite metric directions: lower WER is better, while |
| higher CTEM and TSR are better. The magnitudes indicate a strong |
| monotonic association, so WER is informative, but the spread in the |
| figure shows that it does not fully determine structured-token |
| correctness. |
|
|
| \begin{figure*}[t]\centering |
| \includegraphics[width=0.86\textwidth]{figures/wer_entity_scatter.pdf} |
| \begin{minipage}[t]{0.43\textwidth}\footnotesize |
| \textbf{CTEM vs WER.} Spearman correlation is -0.73: lower WER |
| generally tracks higher entity-level recovery, while substantial spread |
| remains. |
| \end{minipage}\hfill |
| \begin{minipage}[t]{0.43\textwidth}\footnotesize |
| \textbf{TSR vs WER.} Spearman correlation is -0.73: lower WER also |
| tracks full workflow success, with a similar association to CTEM\@. |
| \end{minipage} |
| \caption{WER compared with entity-sensitive metrics across 12 |
| baseline ASR systems, with WER shown on the shared vertical axis. Each |
| point is one baseline ASR system. The negative monotonic correlations show that ordinary |
| transcript accuracy is related to structured-token recovery, while the |
| spread of points shows that WER does not fully determine CTEM or TSR.}\label{fig:wer-entity-scatter} |
| \end{figure*} |
|
|
| \subsection{Performance by Entity |
| Type}\label{performance-by-entity-type} |
|
|
| The evaluation artifacts contain 17,784 model-entity decisions across 12 |
| baseline ASR systems, of which 2,489 are entity failures. These failures |
| are concentrated in punctuation- and boundary-sensitive classes: URLs, |
| commands, email addresses, file paths, and postal addresses account for |
| 1,484 failures, or 59.6\% of all entity failures. Because raw failure |
| counts reflect both entity prevalence and entity difficulty, we treat |
| these counts descriptively; the per-type CTEM results in this section |
| are the main basis for comparing entity difficulty. |
|
|
| Entity-type results show that the hardest values are those whose written |
| forms depend on symbols, separators, and exact token boundaries. Median |
| CTEM is lowest for \texttt{command} (45.0), \texttt{url} (48.4), |
| \texttt{file\_path} (56.9), \texttt{postal\_address} (57.5), and |
| \texttt{email\_address} (65.4). In contrast, highly conventional |
| numeric forms such as \texttt{percentage}, \texttt{measurement}, |
| \texttt{date}, and \texttt{plain\_number} are recovered reliably by |
| most systems. The same contrast appears in the taxonomy groups in |
| Table~\ref{tab:entity-taxonomy}, where numeric and measurement values and |
| language-form entities are high at 97.3 and 95.0 CTEM, while network/web |
| and code/system values trail at 68.7 and 69.4 because they concentrate |
| URLs, commands, file paths, flags, and environment variables. The largest |
| cross-model spreads occur for \texttt{url}, |
| \texttt{command}, \texttt{file\_path}, and |
| \texttt{environment\_variable}, making these types especially useful |
| for differentiating ASR systems beyond WER\@. For the same reason, |
| domain-level aggregates are best read descriptively, since workflow |
| domains differ mainly in their entity mix. Appendix |
| Tables~\ref{tab:appendix-domain-performance} and |
| \ref{tab:appendix-difficulty-performance} provide descriptive domain and |
| difficulty slices for reference. |
|
|
| Table~\ref{tab:error-examples} shows representative cases extracted |
| from the baseline evaluations that illustrate the distinction between |
| \textbf{recoverable formatting variation} and \textbf{unrecoverable |
| evidence loss}. Some transcripts omit canonical formatting while |
| preserving enough information to reconstruct the value. Others destroy |
| essential evidence, such as digits, symbols, separators, or units. |
| VoiceCodeBench counts the former as correct when the value remains |
| uniquely recoverable and the latter as incorrect. |
|
|
| \begin{table*}[t]\centering\footnotesize\begin{tabularx}{\textwidth}{LLLLL} |
| \toprule |
| Entity type & Gold canonical & Transcript evidence & Error class & |
| Recoverable? |
| \\ |
| \midrule |
| \texttt{cli\_flag} & \texttt{-\kern0pt-dry-run} & ``dry run'' & symbol loss & |
| No |
| \\ |
| \texttt{environment\_variable} & \texttt{DATABASE\_URL} & ``database |
| URL'' & underscore/casing loss & No |
| \\ |
| \texttt{phone\_number} & \texttt{212\mbox{-}555\mbox{-}0104} & ``two one two five five |
| five zero one zero four'' & formatting variation & Yes |
| \\ |
| \texttt{file\_path} & \texttt{/var/log/.\kern0pt.\kern0pt./error.log} & ``var log |
| engine x error log'' & path flattening / substitution & No |
| \\ |
| \texttt{currency\_amount} & \texttt{\$7,930.79} & ``seven thousand nine |
| hundred thirty dollars and seventy nine cents'' & formatting variation & |
| Yes |
| \\ |
| \texttt{measurement} & \texttt{500 mg} & ``500 micrograms'' & unit error |
| & No |
| \\ |
| \bottomrule |
| \end{tabularx}\caption{Examples from the baseline evaluations illustrating recoverable and unrecoverable structured-token cases.}\label{tab:error-examples}\end{table*} |
|
|
| \section{Discussion}\label{discussion} |
|
|
| \subsection{Principal Findings}\label{principal-findings} |
|
|
| VoiceCodeBench is designed to measure a production failure mode that |
| is easy to miss with ordinary transcript-level evaluation. The central |
| empirical finding is that WER is useful but incomplete as a predictor of |
| application-facing ASR reliability. Across systems, lower WER generally |
| corresponded to better structured-token recovery, with Spearman |
| correlation $\rho=-0.73$ for CTEM and $\rho=-0.73$ for TSR in |
| Figure~\ref{fig:wer-entity-scatter}. That association means WER remains |
| informative. The spread across systems means it cannot stand in for |
| entity-level or workflow-level correctness. An ASR system can produce a |
| fluent, low-WER transcript while still corrupting the exact structured |
| values that make a workflow usable. The baseline results show that this |
| risk remains present even for strong modern ASR systems. The best TSR in |
| this release was 68.7\%, so nearly one third of recordings still |
| contained at least one unrecovered workflow-critical value. |
|
|
| The system with the lowest WER and highest TSR was Deepgram Nova-3 |
| batch, while the highest CTEM came from ElevenLabs Scribe v2 batch. The |
| weakest structured-token performance appeared for Amazon Transcribe |
| streaming and AssemblyAI Universal-3 Pro streaming. Streaming usually |
| reduced CTEM and TSR relative to paired batch systems, with OpenAI as the |
| exception in this run. These rank differences are practically important. A |
| provider choice that looks best under WER is not necessarily the one |
| that minimizes value-level repair work, and a system that recovers many |
| individual entities can still fail full workflows when errors are spread |
| across recordings. |
|
|
| The most informative failures were not uniformly distributed. The largest |
| performance gaps occurred for URLs, commands, file paths, email |
| addresses, postal addresses, and environment variables. Apparent domain |
| differences largely followed where these entity types appeared. These |
| classes are particularly useful for differentiating systems because they |
| require preservation of symbols, separators, digit sequences, casing |
| cues, or token boundaries rather than only recognition of ordinary words. |
| This concentration is one of the benchmark's central findings. Exact |
| structured-token reliability depends on the written conventions of the |
| entity type, not only on the acoustic difficulty of the surrounding |
| sentence. |
|
|
| \subsection{What Entity-Sensitive Evaluation |
| Reveals}\label{what-entity-sensitive-evaluation-reveals} |
|
|
| WER remains useful for measuring broad transcription quality, but |
| VoiceCodeBench shows why it is incomplete for application-facing |
| speech interfaces. Word-level edit distance can dilute the importance of |
| rare but critical values. A transcript may correctly recognize a long |
| surrounding sentence while losing the one value that the application |
| needs to parse, route, store, compare, or execute. |
|
|
| CTEM and TSR make this failure mode visible. CTEM asks whether each |
| target value remains recoverable from the transcript. TSR asks whether |
| the whole segment is safe for the corresponding workflow. Per-type |
| recovery identifies which value classes are fragile. Reporting these |
| metrics alongside WER allows us to distinguish three cases. These are |
| systems that are broadly accurate and preserve structured values, |
| systems that are broadly accurate but unreliable on structured values, |
| and systems that have both high transcript error and high |
| structured-token error. |
|
|
| This distinction changes how ASR results can be used by product and |
| platform teams. WER is appropriate for monitoring general transcript |
| quality, but CTEM estimates value-level correction burden, TSR estimates |
| the fraction of recordings that can flow through automation without |
| human or programmatic repair, and per-type recovery points to targeted |
| mitigations. A team building support routing, form filling, developer |
| tooling, healthcare administration, or billing workflows may therefore |
| choose different ASR settings, add confirmation for only certain entity |
| types, or reject an otherwise strong model if its failures cluster in |
| high-cost values. The correlation analysis reinforces this point. Lower |
| WER is associated with higher CTEM and TSR across systems, but the |
| relationship is not deterministic, so aggregate transcript accuracy can |
| hide weaknesses in specific entity classes. |
|
|
| \subsection{Punctuation and Normalization as |
| Correctness}\label{punctuation-and-normalization-as-correctness} |
|
|
| A recurring theme in VoiceCodeBench is that punctuation, formatting, |
| and normalization are not merely cosmetic. In ordinary transcript |
| readability tasks, punctuation and capitalization may be treated as |
| presentation details or post-processing targets. In structured workflow |
| tokens, they can define the value itself. Dashes, dots, slashes, |
| underscores, equals signs, leading zeros, decimal points, units, and |
| token boundaries may determine whether a value is parseable and correct. |
|
|
| This is why the benchmark distinguishes between acoustic and canonical |
| forms. The acoustic form captures what the speaker said; the canonical |
| form captures what the application needs. The relevant question is not |
| whether the transcript exactly matches a reference sentence, but whether |
| it preserves enough evidence to recover the intended canonical value. |
| For example, a phone number may remain recoverable even if hyphens are |
| omitted, because the digit sequence is intact. By contrast, a file path |
| or email address may become unrecoverable if slashes, dots, or |
| separators are lost. |
|
|
| The baseline examples in Table~\ref{tab:error-examples} support this |
| distinction. Failures such as URL domain corruption, command |
| substitution, path segment loss, email-domain corruption, suite/unit |
| confusion, and environment-variable word-form substitution are not |
| merely formatting imperfections. They change the value available to |
| downstream software. This suggests that ASR evaluation for workflow use |
| should treat punctuation-sensitive and normalization-sensitive values as |
| first-class targets rather than as secondary presentation features. |
|
|
| \subsection{Intended Use}\label{intended-use} |
|
|
| VoiceCodeBench is intended as a diagnostic benchmark for |
| structured-token reliability in ASR systems. It supports provider |
| comparison, regression tracking, per-entity risk analysis, and decisions |
| about application-specific safeguards. Strong aggregate performance is |
| not sufficient when weak recovery for entities such as |
| \texttt{file\_path}, \texttt{email\_address}, or |
| \texttt{currency\_amount} may still require validation, confirmation |
| prompts, constrained decoding, or downstream repair before production |
| use. |
|
|
| The practical implication for applications like \textbf{voice agents} is that |
| transcript readability and low WER are insufficient deployment gates |
| when ASR output feeds tool calls, form fills, database writes, routing |
| logic, or command execution. Agent pipelines should evaluate whether the |
| specific values they act on remain recoverable, then pair ASR selection |
| with product controls such as confirmation prompts, typed validation, |
| constrained inputs, or human review for high-risk entity classes. |
|
|
| The benchmark is not a general measure of ASR quality, a training |
| corpus, a fine-tuning set, or a hidden leaderboard. It does not cover all |
| speech styles, languages, acoustic environments, or conversational |
| settings. Because labels are public, reports should disclose model |
| versions, evaluation dates, inference settings, and any |
| benchmark-specific prompting, post-processing, fine-tuning, or |
| canonicalization. Its best use is comparative and diagnostic, identifying |
| which systems preserve exact structured values, which entity types or |
| workflow slices are fragile, and which errors remain recoverable from |
| transcript evidence. |
|
|
| \subsection{Limitations}\label{limitations} |
|
|
| VoiceCodeBench is intentionally focused, and its results should be |
| interpreted within that scope. First, the benchmark is English-only. |
| Structured-token dictation is common in many languages, but punctuation |
| conventions, spelling practices, number formats, address formats, and |
| spoken symbol conventions vary across languages. Multilingual evaluation |
| will require additional design work rather than direct translation. |
|
|
| Second, the benchmark focuses on compact workplace-style dictation. It |
| does not attempt to cover all ASR conditions, including casual |
| conversation, meetings, overlapping speech, broadcast audio, voice |
| search, noisy field recordings, or long-form dictation. This narrowness |
| is deliberate because the benchmark targets a specific workflow failure |
| mode. |
| However, results should not be generalized to all speech-recognition use |
| cases. |
|
|
| Third, the scenarios and values are synthetic, even though the audio is |
| human-recorded. Synthetic content allows controlled coverage and safer |
| public release, but it may not capture every distributional property of |
| real support calls, developer conversations, clinical documentation, |
| logistics workflows, or financial operations. Some generated scenarios |
| may be cleaner or more compact than naturally occurring speech. |
|
|
| Fourth, exact structured-token scoring requires policy choices and |
| currently uses LLM-assisted recoverability verification. VoiceCodeBench |
| counts an entity as correct if the transcript contains enough evidence |
| to recover the canonical value exactly, even when that evidence appears |
| in spoken or partially normalized form. This is more flexible than |
| literal string match, but it requires careful handling of borderline |
| cases and introduces dependence on the verifier model and prompt. Some |
| values are clearly recoverable despite formatting variation; others |
| become ambiguous when symbols, separators, or units are omitted. The |
| benchmark mitigates this by documenting scoring rules, the verifier |
| artifact \texttt{openai\_gpt\_5\_5\_v1}, its model \texttt{gpt-5.5}, the |
| prompt, the strict JSON response schema, the absence of a temperature override |
| for GPT-5-family verifier models, |
| per-model verifier JSON files with evidence spans and reasons, and final |
| scoring files. A human audit of 200 entity decisions sampled across ASR |
| models and entity types found 100\% agreement with the verifier decisions. |
| Future versions should expand deterministic scoring for entity types where |
| exact canonicalization can be implemented reliably, but some edge cases will |
| remain debatable. |
|
|
| Fifth, commercial ASR systems change over time. Provider APIs may update |
| models, defaults, formatting behavior, or streaming endpointing without |
| preserving old behavior. For that reason, results should be reported |
| with evaluation dates and configuration details. VoiceCodeBench is |
| most useful when treated as a repeatable evaluation protocol rather than |
| a one-time static ranking. |
|
|
| \subsection{Ethical and Privacy |
| Considerations}\label{ethical-and-privacy-considerations} |
|
|
| The primary privacy risk is speaker identifiability. Although scripts |
| and structured values are synthetic, released audio can contain voice |
| characteristics that identify or profile speakers. Contributors are paid |
| and consent to dataset use and release, speaker metadata is limited to |
| broad optional categories and anonymous identifiers, and sensitive or |
| routable structured values are avoided by construction. The dataset is |
| intended for ASR evaluation, not for speaker identification, biometric |
| modeling, demographic profiling, or training production |
| speech-recognition systems. |
|
|
| \subsection{Future Work}\label{future-work} |
|
|
| Future extensions could broaden VoiceCodeBench's coverage. A |
| multilingual version would test structured-token recovery across |
| languages, number systems, spelling conventions, and punctuation |
| practices. Expanded speaker coverage would allow more robust analysis by |
| accent, region, device, and recording condition, provided such metadata |
| is collected and reported responsibly. Additional acoustic conditions, |
| such as noise, reverberation, compression, and telephony codecs, would |
| help measure whether structured-token recovery degrades under realistic |
| deployment conditions. |
|
|
| A contextual extension would also be valuable. The current benchmark |
| measures raw-audio-only ASR behavior. A future version could compare |
| this setting against ASR systems given domain labels, candidate entity |
| lists, custom vocabulary, grammars, or application-specific |
| canonicalizers. This would help quantify how much context improves exact |
| structured-token recovery and which entity types benefit most from |
| additional information. |
|
|
| \subsection{Conclusion}\label{conclusion} |
|
|
| VoiceCodeBench focuses attention on a practical ASR requirement that |
| broad transcript metrics can obscure. ASR systems must preserve the exact |
| structured values that downstream workflows depend on. Its contribution is not |
| another general ASR corpus, but a test of whether raw-audio ASR output |
| preserves enough evidence to recover identifiers, paths, commands, |
| measurements, addresses, and other written values exactly. By combining |
| entity-first dataset construction, acoustic/canonical annotations, |
| raw-audio-only evaluation, and entity-sensitive metrics, the benchmark |
| connects transcript evaluation to workflow risk. |
|
|
| The baseline results show why this distinction matters. WER is |
| informative, but it does not determine CTEM or TSR; the strongest |
| systems still leave many recordings with at least one unrecovered |
| workflow-critical value; and the hardest cases concentrate in |
| punctuation-, separator-, and boundary-sensitive entity types. Reporting |
| WER alongside CTEM, TSR, and per-type recovery therefore gives ASR |
| developers and application teams a more actionable view of quality: |
| which systems produce readable transcripts, which preserve exact values, |
| and where additional safeguards are needed before speech output can be |
| trusted as software input. |
|
|
| \section{Data Availability}\label{data-availability} |
|
|
| VoiceCodeBench is released as a public, test-only benchmark. It is not |
| intended for model training, fine-tuning, or post-training. The dataset |
| and evaluation scripts are made available through the Hugging Face |
| dataset repository at |
| \url{https://huggingface.co/datasets/besimple-ai/voice-code-bench}. |
|
|
| VoiceCodeBench provides the audio, reference transcripts, entity |
| annotations, metadata, scoring scripts, baseline outputs, aggregate |
| result tables, and documentation needed to reproduce evaluation. |
| Because labels are public, VoiceCodeBench should be treated as a |
| transparent diagnostic benchmark rather than a hidden leaderboard. |
| Reported results should disclose the model version, evaluation date, |
| inference settings, and any benchmark-specific prompting, fine-tuning, |
| post-processing, or canonicalization. |
|
|
| \section{References}\label{references} |
|
|
| Ardila, R., Branson, M., Davis, K., Henretty, M., Kohler, M., Meyer, J., |
| Morais, R., Saunders, L., Tyers, F., and Weber, G. 2020. Common Voice: A |
| Massively-Multilingual Speech Corpus. In \emph{Proceedings of the 12th |
| Language Resources and Evaluation Conference (LREC 2020)}, 4218--4222. |
|
|
| Arkko, J., Cotton, M., and Vegoda, L. 2010. IPv4 Address Blocks Reserved |
| for Documentation. RFC 5737. |
|
|
| Bender, E. M., and Friedman, B. 2018. Data Statements for Natural |
| Language Processing: Toward Mitigating System Bias and Enabling Better |
| Science.\ \emph{Transactions of the Association for Computational |
| Linguistics}, 6:587--604. |
|
|
| Carletta, J. 2007. Unleashing the Killer Corpus: Experiences in Creating |
| the Multi-Everything AMI Meeting Corpus.\ \emph{Language Resources and |
| Evaluation}, 41(2):181--190. |
|
|
| Chen, G., Chai, S., Wang, G., Du, J., Zhang, W.-Q., Weng, C., Su, D., |
| Povey, D., Trmal, J., Zhang, J., et al. 2021. GigaSpeech: An Evolving, |
| Multi-Domain ASR Corpus with 10,000 Hours of Transcribed Audio. In |
| \emph{Proceedings of Interspeech 2021}. |
|
|
| Conneau, A., Ma, M., Khanuja, S., Zhang, Y., Axelrod, V., Dalmia, S., |
| Riesa, J., Rivera, C., and Bapna, A. 2022. FLEURS\@: Few-shot Learning |
| Evaluation of Universal Representations of Speech. In \emph{Proceedings |
| of the IEEE Spoken Language Technology Workshop (SLT)}, 798--805. |
|
|
| Del Rio, M., Delworth, N., Westerman, R., Huang, M., Bhandari, N., |
| Palakapilly, J., McNamara, Q., Dong, J., Zelasko, P., and Jette, M. |
| 2021. Earnings-21: A Practical Benchmark for ASR in the Wild. In |
| \emph{Proceedings of Interspeech 2021}. |
|
|
| Eastlake, D., and Panitz, A. 1999. Reserved Top Level DNS Names. RFC |
| 2606. |
|
|
| Gebru, T., Morgenstern, J., Vecchione, B., Vaughan, J. W., Wallach, H., |
| Daum\'e III, H., and Crawford, K. 2021. Datasheets for Datasets. |
| \emph{Communications of the ACM}, 64(12):86--92. |
|
|
| Kim, S., Arora, A., Le, D., Yeh, C.-F., Fuegen, C., Kalinli, O., and |
| Seltzer, M. L. 2021. Semantic Distance: A New Metric for ASR Performance |
| Analysis Towards Spoken Language Understanding. In \emph{Proceedings of |
| Interspeech 2021}. |
|
|
| Meister, A., Novikov, M., Karpov, N., Bakhturina, E., Lavrukhin, V., and |
| Ginsburg, B. 2023. LibriSpeech-PC\@: Benchmark for Evaluation of |
| Punctuation and Capitalization Capabilities of End-to-End ASR Models. |
| arXiv:2310.02943. |
|
|
| National Institute of Standards and Technology. 2021. OpenASR21 |
| Challenge Evaluation Plan. |
|
|
| North American Numbering Plan Administrator. 555 Line Numbers. |
|
|
| Panayotov, V., Chen, G., Povey, D., and Khudanpur, S. 2015. LibriSpeech: |
| An ASR Corpus Based on Public Domain Audio Books. In \emph{Proceedings |
| of the IEEE International Conference on Acoustics, Speech and Signal |
| Processing (ICASSP 2015)}, 5206--5210. |
|
|
| Roy, S. 2021. Semantic-WER\@: A Unified Metric for the Evaluation of ASR |
| Transcript for End Usability.\ arXiv:2106.02016. |
|
|
| Szymanski, P., Augustyniak, L., Morzy, M., Szymczak, A., Surdyk, K., and |
| Zelasko, P. 2023. Why Aren't We NER Yet? Artifacts of ASR Errors in |
| Named Entity Recognition in Spontaneous Speech Transcripts. In |
| \emph{Proceedings of the 61st Annual Meeting of the Association for |
| Computational Linguistics (Volume 1: Long Papers)}, 1746--1761. |
|
|
| Tan, S., Behre, P., Kibre, N., Alphonso, I., and Chang, S. 2022. |
| Four-in-One: A Joint Approach to Inverse Text Normalization, |
| Punctuation, Capitalization, and Disfluency for Automatic Speech |
| Recognition.\ arXiv:2210.15063. |
|
|
| Wang, Y.-Y., Acero, A., and Chelba, C. 2003. Is Word Error Rate a Good |
| Indicator for Spoken Language Understanding Accuracy? In |
| \emph{Proceedings of the IEEE Workshop on Automatic Speech Recognition |
| and Understanding (ASRU 2003)}, 577--582. |
|
|
| Wang, H., Ma, L., Guo, D., Wang, X., Xie, L., Xu, J., and Lin, J. 2025. |
| ContextASR-Bench: A Massive Contextual Speech Recognition Benchmark. |
| arXiv:2507.05727. |
|
|
| \appendix |
|
|
| \section{Appendix: Domain and Difficulty |
| Slices}\label{appendix-domain-difficulty-slices} |
|
|
| Tables~\ref{tab:appendix-domain-performance} and |
| \ref{tab:appendix-difficulty-performance} report descriptive CTEM and |
| TSR slices averaged across the 12 released baseline systems. These |
| tables are intended as secondary context rather than provider rankings: |
| domains and difficulty bands differ in entity mix and entity load. |
|
|
| \begin{table}[!htbp]\centering\scriptsize |
| \caption{Performance by workflow domain, averaged across the 12 baseline systems.}\label{tab:appendix-domain-performance} |
| \begin{tabular}{lrrrr} |
| \toprule |
| Domain & Rec. & Ent. & CTEM & TSR |
| \\ |
| \midrule |
| Contact/routing & 45 & 213 & 83.5 & 42.4 |
| \\ |
| Technical/IT/dev & 55 & 260 & 70.5 & 25.8 |
| \\ |
| Retail/logistics & 45 & 214 & 85.3 & 45.9 |
| \\ |
| Finance/billing & 40 & 190 & 84.9 & 49.0 |
| \\ |
| Healthcare/admin & 35 & 167 & 96.7 & 85.2 |
| \\ |
| Legal/ins./gov. & 35 & 167 & 95.0 & 80.2 |
| \\ |
| Education/workplace & 25 & 119 & 92.4 & 71.3 |
| \\ |
| Dense mixed stress & 20 & 152 & 91.8 & 55.4 |
| \\ |
| \bottomrule |
| \end{tabular} |
| \end{table} |
|
|
| \begin{table}[!htbp]\centering\scriptsize |
| \caption{Performance by difficulty band, averaged across the 12 baseline systems.}\label{tab:appendix-difficulty-performance} |
| \begin{tabular}{lrrrr} |
| \toprule |
| Difficulty & Rec. & Ent. & CTEM & TSR |
| \\ |
| \midrule |
| Light & 30 & 90 & 85.8 & 66.9 |
| \\ |
| Standard & 114 & 456 & 83.6 & 54.5 |
| \\ |
| Dense & 93 & 479 & 86.1 & 53.6 |
| \\ |
| Stress & 63 & 457 & 88.3 & 45.0 |
| \\ |
| \bottomrule |
| \end{tabular} |
| \end{table} |
|
|
| \section{Appendix: Generation Tooling}\label{appendix-generation-tooling} |
| \begingroup\raggedright |
| Transcript and entity generation used OpenAI Codex as a repository-aware |
| LLM agent under two repository-local skills: |
| \texttt{voice-code-bench-}\allowbreak\texttt{generate-transcript} for |
| single segments and |
| \texttt{voice-code-bench-}\allowbreak\texttt{generate-transcript-loop} |
| for dataset-wide iteration. These workflows preserved each segment's |
| domain, scenario, difficulty, entity constraints, acoustic and canonical |
| annotations, and template/acoustic/canonical transcript layers in |
| \texttt{data/metadata.jsonl}, while checking recoverability and taxonomy |
| alignment before acceptance. |
| \par\endgroup |
|
|
| \clearpage |
| \onecolumn |
| \section{Appendix: Verifier Prompt}\label{appendix-verifier-prompt} |
|
|
| The verifier artifact \texttt{openai\_gpt\_5\_5\_v1} uses the following |
| system prompt, line-wrapped here for print. The user message supplies |
| the datapoint ID, target entities, and ASR transcript as JSON; ``STT |
| transcript'' denotes the ASR output being verified. |
|
|
| \begingroup |
| \scriptsize |
| \begin{verbatim} |
| You verify whether each gold Voice Code Bench entity is present in a |
| raw speech-to-text transcript. |
| |
| Return only valid JSON matching the provided schema. |
| |
| Entity types: |
| - email_address: full email addresses, including dots, hyphens, |
| underscores, plus tags, and spoken at/dot separators. |
| - phone_number: dialable phone numbers. The canonical form uses |
| XXX-XXX-XXXX for US numbers in this dataset. |
| - phone_extension: phone extension values such as ext74 or ext4821. |
| - person_or_team_name: unambiguous public, team, organization, or |
| routing names whose written form is recoverable from the transcript. |
| - postal_address: mailing addresses, suites, floors, units, cities, |
| state abbreviations, and ZIP-like values. |
| - url: web URLs and hostnames, including subdomains, paths, and query |
| strings. |
| - ip_address: IPv4 addresses in dotted decimal form. |
| - port_number: network port numbers. |
| - command: literal CLI commands or command snippets whose exact tokens |
| matter. |
| - cli_flag: command-line flags such as --dry-run, --config, or -k. |
| - file_path: file paths, directory paths, filenames, hidden files, and |
| extensions. |
| - environment_variable: environment variable names such as DATABASE_URL |
| or NODE_ENV. |
| - code_symbol: function names, class names, package names, branch names, |
| config keys, and identifiers. |
| - version: software, firmware, API, schema, or model versions. |
| - reference_id: cases, tickets, claims, invoices, appointments, |
| confirmations, tracking numbers, and other operational IDs. |
| - product_code: SKUs, serials, model numbers, part numbers, lot numbers, |
| and device identifiers. |
| - account_or_record_number: account numbers, masked account tails, |
| medical record numbers, member IDs, and record locators. |
| - currency_amount: monetary amounts with an explicit currency. |
| - percentage: percentages, rates, APRs, tax rates, discounts, and |
| allocation percentages. |
| - measurement: numeric values with units, including dosage, weight, |
| length, volume, temperature, pressure, duration, and lab values. |
| - plain_number: exact standalone numbers without a unit or currency. |
| - date: calendar dates where the exact date matters. |
| - time: appointment times, deadlines, time windows, and time zones. |
| - acronym_or_initialism: spoken or letter-by-letter acronyms and |
| initialisms, including conventional punctuation. |
| - spelled_sequence: values explicitly spelled letter-by-letter, |
| including names and mixed letter/digit sequences. |
| - domain_term: specialized vocabulary that matters for task success and |
| is not covered by a more structured type. |
| |
| Rules: |
| - Use only the STT transcript. |
| - Check every target entity independently and return one result for each |
| target_index. |
| - Use the target acoustic field as the expected spoken form and the |
| canonical field as the exact value to recover. |
| - Mark present true only when the transcript contains enough evidence to |
| recover that exact canonical value. |
| - Accept casing, punctuation, spacing, and formatting differences only |
| when the same value remains recoverable. |
| - Reject wrong, missing, extra, or substituted letters, digits, |
| separators, units, dates, times, amounts, or words that change the |
| target value. |
| - Copy target_index, type, and canonical exactly from the target entity |
| input. |
| - Set present true only when the target is supported by the transcript. |
| - Include an evidence field with the exact transcript substring that |
| supports the decision. For absent entities, use the closest corrupted |
| substring when available, otherwise use an empty string. |
| - Include a short reason explaining the present/absent decision. |
| - Do not return entities that are not listed as target entities. |
| - Do not explain outside the JSON. |
| \end{verbatim} |
| \endgroup |
|
|
| \end{document} |
|
|
