Datasets:
Download README.md from SebastianBodza/decider-audio-when2call: direct link, hf CLI and curl.
- Browser
- Download file 17.5 kB
-
https://huggingface.co/datasets/SebastianBodza/decider-audio-when2call/resolve/main/README.md
- Command line
-
hf download hf://datasets/SebastianBodza/decider-audio-when2call/README.md
-
curl -L -o README.md https://huggingface.co/datasets/SebastianBodza/decider-audio-when2call/resolve/main/README.md
pretty_name: Decider Audio – When2Call (spoken)
language:
- en
license: other
license_name: mixed-see-card
task_categories:
- audio-classification
- text-classification
size_categories:
- 10K<n<100K
tags:
- audio
- tts
- synthetic
- tool-calling
- function-calling
- when2call
- decision
- decider-audio
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
Decider Audio – When2Call (spoken)
Part of the Decider Audio collection.
The model hears a user question. It reads the assistant's tool definitions and the candidate replies as text. Then it must decide which reply the assistant should give. The reply can call a tool, ask a follow-up question, or say that the assistant cannot help. The words alone can solve the task, so the challenge is to understand them from speech and to match them with the tools.
12,622 rows from NVIDIA's When2Call, about 21.67 hours of audio over all rows. 11,269 distinct spoken questions were rendered by 6 text-to-speech models.
| Split | Source | Rows | Spoken clips | Distinct voices | Hours | Re-rendered | Failing checks |
|---|---|---|---|---|---|---|---|
train |
When2Call train_pref (2 options) |
8,970 | 8,970 | 204 | 14.81 | 199 | 11 |
test |
When2Call test MCQ (4 options, official evaluation set) |
3,652 | 2,299 | 202 | 6.87 | 56 | 1 |
No speaker appears in more than one split. The test rows are spoken by test-split voices only.
Task format
| Input | audio (the spoken question) + instruction + tools + choices as text |
| Target text | answer, an option key such as option_3 |
| Resolved label | answer_label, which is tool_call, request_for_info or cannot_answer in test and chosen in train |
Here is a suggested prompt. The model sees the audio at the marked position.
Available tools:
[{"name": "...", "description": "...", "parameters": {...}}, ...]
User question:
<audio>
Which response should the assistant give to the user's question, given the available tools?
option_0: ...
option_1: ...
Answer with the option key.
spoken_text is the original question. It is included for quality control and ASR baselines. Do not give it to a model as input when training an audio model.
How it is made
- Text.
test= When2Calltest/when2call_test_mcq.jsonl(3,652 rows, the official multiple-choice evaluation, must not be used for training). It has four candidate replies in the fixed orderdirect,tool_call,request_for_info,cannot_answer, anddirectis never correct.train=train/when2call_train_pref.jsonl(9,000 rows). It has the chosen and the rejected reply in a seeded shuffle per row. Questions, options and answer keys are identical to the Decision Index projection of When2Call. The SFT training file (single replies, no options) is not used. - Content-only speech. Only the user question is spoken. Tools, instruction, options and answer never enter the audio.
- One clip per test question. When2Call asks many test questions twice or more. Once with the tool it needs (
tool_call) and once with that tool removed (cannot_answer), sometimes also asrequest_for_info. These rows share one clip (spoken_unit,spoken_unit_rows> 1), so the audio is identical and only the tools decide. 2,640testrows share their clip with another row. 2,299 distinct test questions were spoken.trainrows are spoken one by one. No row was merged. - Model and voice assignment by question type. The six models are not equally good at everything. A pilot test rendered 43 hard questions with every model and the same voice. On plain questions all models are equally good, but lists, ID tokens, URLs, handles and long numbers separate them. Each question is classified (
speech_category, the first matching type in the order list, URL, ID token, handle, long number) and gets the best-ranked model of its type that still has room. Types are assigned from the smallest to the largest, and plain questions fill the remaining slots least-used-first. Every model still renders the same number of clips (±1) per split. The assignment does not look at the answer. Lists and ID tokens are spoken without list brackets and without the quotes around values (tts_text, e.g. "Split the list ['a', 'b']" becomes "Split the list a, b"). This made lists clearly easier for every model except OmniVoice. The voice is picked by a seeded hash from the pool of the matching split.
| Question type | Spoken as | Ranking from the pilot | Clips | chatterbox |
omni |
pocket |
qwen |
sopro |
supertonic |
|---|---|---|---|---|---|---|---|---|---|
| handle | original | supertonic > omni > qwen > sopro > chatterbox > pocket | 218 | 1 | 0 | 0 | 1 | 0 | 216 |
| url | original | supertonic > omni > chatterbox > qwen > sopro > pocket | 372 | 1 | 3 | 1 | 3 | 1 | 363 |
| list | cleaned | sopro > pocket > chatterbox > qwen > supertonic > omni | 518 | 0 | 0 | 0 | 0 | 518 | 0 |
| id token | cleaned | qwen > omni > pocket > chatterbox > sopro > supertonic | 625 | 0 | 0 | 0 | 625 | 0 | 0 |
| long number | original | omni > pocket > qwen > supertonic > chatterbox > sopro | 2,499 | 0 | 1,830 | 669 | 0 | 0 | 0 |
| plain | original | balanced | 6,988 | 1,878 | 0 | 1,209 | 1,253 | 1,360 | 1,288 |
| non-English | own language | OmniVoice only | 49 | 0 | 49 | 0 | 0 | 0 | 0 |
- Non-English questions. 49 spoken questions (85 rows) are not English. They are 11
mixed, 11zh, 6ko, 6id, 3vi, 3es, 3th, 2hi, 2pt, 1hi-latnand 1de. They were found by script (Han, kana, Hangul, Thai, Devanagari), a language detector and a keyword search, and reviewed by hand. They are always spoken by OmniVoice, the only multilingual model, in their language (language-agnostic mode formixed). It uses the same English reference voices (cross-lingual cloning). They were moved to OmniVoice after the balanced assignment, so the other questions keep their models. They are checked withSystran/faster-whisper-large-v3and the character error rate. - Voices. Voice-cloning models use English reference clips from a private voice collection (not redistributed). Speakers are assigned to
train,validationandtestbefore any voice is chosen, and every row uses only voices of its own split, so test speakers never occur in training. Preset-voice models use the presets of their split's partition. 406 distinct voices appear in total. - Audio. 24 kHz mono, 16-bit FLAC. Clips whose peak exceeded full scale were scaled to 0.99 (
audio_gain). Flat-topped (clipped) renders were rejected instead. - Re-rendering, never removal. Every render was checked. A render fails if the ASR word error rate is above 0.5, if there are more than 30 characters per second, or if 3 or more extra words appear that the question does not contain (10 % of its words for questions longer than 30 words). For non-English questions the error rate is the character error rate and the extra-word check does not apply. A failed question was rendered again with a different voice and seed, using the same model (
render_attempt> 0, up to 5 attempts). 255 rows carry such a re-render. 12 rows still failed after the last attempt and keep their best render withpasses_checks = false. A second repair pass rendered 137 clips again. 42 of them still failed a check and 95 passed the word check but leave out part of their text. A clip got up to 10 more attempts on its own model. Clips on Supertonic and clips that a model could not render at all then moved to another model for up to 3 more. 12 clips now come from another model, whichtts_modelshows. A clip now also passes when its only problem is the word error rate or extra words and its transcript sounds like the text. This check compares phonemes after numbers and units are spelled out, so "topup" and "top up" or "1960" and "nineteen sixty" match. A clip that leaves out part of its text does not pass this way. A truncated clip is replaced only when the new render leaves out less.
TTS models
| TTS model | Checkpoint | Voice mode | Rows | Distinct voices | Hours | Mean WER | Exact | WER ≤ 0.2 | Re-rendered |
|---|---|---|---|---|---|---|---|---|---|
chatterbox |
ResembleAI/chatterbox | clone | 2,103 | 376 | 2.85 | 0.0407 | 70.0% | 95.0% | 14 |
omni |
k2-fsa/OmniVoice | clone | 2,125 | 377 | 4.48 | 0.0394 | 59.2% | 97.0% | 29 |
pocket |
kyutai/pocket-tts | clone | 2,109 | 370 | 3.50 | 0.0390 | 65.7% | 96.5% | 49 |
qwen |
Qwen/Qwen3-TTS-12Hz-0.6B-Base | clone | 2,107 | 371 | 3.63 | 0.0558 | 59.3% | 91.6% | 87 |
sopro |
samuel-vitorino/sopro-v2-turbo | clone | 2,110 | 371 | 3.70 | 0.0334 | 72.7% | 96.1% | 32 |
supertonic |
Supertone/supertonic-3 | preset | 2,068 | 7 | 3.52 | 0.0556 | 60.3% | 92.9% | 44 |
Counts are rows. A shared test clip counts once per row. OmniVoice, Chatterbox and Qwen are the original PyTorch checkpoints.
Checks
Every clip was transcribed with Systran/faster-whisper-base.en and compared with spoken_text after lower-casing, removing punctuation and spelling out digits.
- Mean error rate over rows 0.0439 (word error rate, and character error rate for the 85 non-English rows). 8,144 rows (64.5%) transcribe exactly. The 90th percentile is 0.143 and the 99th is 0.385.
- 94.9% of the rows have WER ≤ 0.2. Questions contain IDs, handles, URLs and file names ("@jane_doe", "project ID badae") that TTS reads in its own way and ASR spells differently. Part of the remaining error is in such tokens. Filter on
asr_werorpasses_checksfor a cleaner subset.
ASR agreement shows that the words were spoken. It does not show that the voice sounds natural or that it resembles the reference voice. Only a pilot sample was reviewed by listening. Voice similarity differs between models.
Rows left out
The dataset follows When2Call row by row with one exception. 30 train questions about Sudoku spell out the whole board ([[5,3,'.','.',7,...],...], '53..7....'). Read aloud, that is a minute of digits and dots that no model rendered intelligibly in five attempts, so these rows are not included. Sudoku questions without a board are kept. The source lines of the left-out train rows are 48, 708, 1179, 1225, 1237, 1459, 1781, 2241, 2341, 2768, 2885, 3097, 3351, 3444, 3459, 3531, 3840, 3850, 4514, 4528, 4816, 5074, 5628, 6494, 6762, 6818, 8352, 8436, 8519 and 8832.
Peculiarities of the source
- Prompt-like questions. 3
testrows have a long system prompt ("You are MemGPT, …") as their question and 15 rows contain an artifact "Revised question: …" that repeats the question. They are spoken in full, which gives the longest clips (41 rows have more than 400 characters). - Repeated questions. Besides the shared test clips, 657 questions occur more than once inside
train(with different tools or replies), and 2trainrows have a question that also occurs intest. To drop those, use this code.
import re
from datasets import load_dataset
def key(text):
return " ".join(re.findall(r"[a-z0-9']+", text.lower().replace("-", " ")))
train = load_dataset("SebastianBodza/decider-audio-when2call", split="train")
test = load_dataset("SebastianBodza/decider-audio-when2call", split="test")
test_questions = {key(t) for t in test["spoken_text"]}
train_clean = train.filter(lambda row: key(row["spoken_text"]) not in test_questions)
# optionally also keep only renders that passed every check
train_clean = train_clean.filter(lambda row: row["passes_checks"])
Fields
| Field | Type | Description |
|---|---|---|
id |
string | when2call-<split>-<index>, index = line of the source file |
audio |
Audio, 24 kHz mono FLAC | Spoken question |
instruction |
string | Task text |
tools |
list of strings | Tool definitions as in the source, one JSON object per entry (may be empty) |
choices |
list of strings | Candidate replies. choices[i] belongs to option key option_i. Tool calls are JSON (test) or <TOOLCALL>[...]</TOOLCALL> (train), as in the source |
choice_labels |
list of strings | What each candidate is. Values are direct, tool_call, request_for_info, cannot_answer (test) and chosen, rejected (train) |
answer |
string | Option key of the correct reply |
answer_index |
int | Index of the correct reply in choices |
answer_label |
string | choice_labels[answer_index] |
spoken_text |
string | The user question as in the source |
tts_text |
string | Text the TTS model got. This is spoken_text, or for lists and ID tokens spoken_text without list brackets and delimiter quotes |
speech_category |
string | Question type that chose the model (list, url, id_token, handle, long_number or plain) |
spoken_language |
string | en, a language code (zh, ko, th, hi, id, vi, es, pt, de), hi-latn (Hindi written in Latin script) or mixed (English with words or a passage in another script) |
source_dataset, source_id, source_split |
string | Provenance (When2Call:<split>:<line>) |
source_uuid, source_subset, target_tool |
string | Test only. When2Call uuid, the BFCL subset it was built from, and the target tool (JSON) |
spoken_unit, spoken_unit_rows |
string, int | Id of the spoken clip and the number of rows that share it |
tts_model |
string | Project key of the TTS model |
tts_repo, tts_revision, tts_sdk_revision |
string | Checkpoint, pinned revision and runtime revision |
voice_id |
string | Reference-voice id (clone models) or preset name |
voice_kind |
clone or preset |
How the voice was chosen |
voice_split |
string | Split the voice belongs to. Always equal to the row's split |
tts_settings |
JSON string | Generation settings |
tts_seed |
int | Seed of the generation call |
duration_seconds, n_segments |
float, int | Length of the clip. Questions above 400 characters are split into ordered segments joined with 180 ms of silence |
audio_gain |
float | 1.0 unless the clip overshot full scale and was scaled down (823 rows) |
asr_text, asr_wer, asr_metric, asr_model |
string, float, string, string | Round-trip check, see above. asr_metric says whether asr_wer is a word (wer) or character error rate (cer, non-English rows) |
render_attempt |
int | 0 for the first render, higher if the question was rendered again after a failed check |
passes_checks |
bool | False if even the last attempt failed a check |
Clips are 0.52–303.71 s long (mean 6.18 s per row).
Limitations
- Synthetic speech only. All audio comes from TTS models, there is no human speech. Expect a domain gap to real users, accents, noise and telephone channels.
- Train and test differ in format.
trainhas two options (chosen vs. rejected reply, no category labels),testfour (one per category). A model trained only ontrainnever sees four options. This is the source's design. - Answer distribution follows the source.
testhas 1,295tool_call, 1,062request_for_infoand 1,295cannot_answer. - Questions were written as text. Many look like typed requests (IDs, quotes, URLs, JSON-like values) rather than natural speech.
- Mostly English. 85 rows are in other languages or mixed. All are spoken by OmniVoice with English reference voices, so they carry an English accent of varying strength. Speaker attributes (gender, age, accent) are not annotated.
License and attribution
The dataset mixes several upstream terms. Check them before any use beyond private research.
- Text. When2Call, NVIDIA, CC BY 4.0 (paper, data). Its test questions and tools derive from BFCL v2 Live.
- Speech from OmniVoice. The upstream weights are reported as CC BY-NC 4.0 (non-commercial). Verify on
k2-fsa/OmniVoice. Thetts_repocolumn shows which checkpoint rendered a clip. - Other generators. Qwen3-TTS Apache-2.0, Sopro Apache-2.0, Supertonic OpenRAIL (use restrictions apply), Pocket TTS CC BY 4.0 (attribution required, and Kyutai's PyTorch weights are gated behind their terms), Chatterbox MIT.
- Reference voices. Taken from a private collection and not redistributed.
Because of the non-commercial and use-restricted components the dataset is marked license: other. Do not use it commercially before the terms above are resolved.
@inproceedings{ross-etal-2025-when2call,
title = {When2Call: When (not) to Call Tools},
author = {Ross, Hayley and Mahabaleshwarkar, Ameya Sunil and Suhara, Yoshi},
booktitle = {Proceedings of NAACL 2025},
year = {2025},
url = {https://aclanthology.org/2025.naacl-long.174/}
}