DoorDash-Audio-Benchmarks

A benchmark dataset of human speech recordings for evaluating speech and speech-to-speech model performance.

Use restriction — read before you download

This dataset is source-available, not open source. You may use, share, and adapt it under the terms of LICENSE, with one key limitation:

You may not use this dataset to create voiceprints or other biometric identifiers, to identify or re-identify any individual whose voice appears in it, or to locate or contact those individuals. See Section 5 of the license.

The license also requires attribution and passes this restriction through to anyone you share the data (or derived data) with.

What's in it

• 2,037 recordings / 60.1 minutes of spoken restaurant-ordering turns — 1,293 human-voice recordings totalling 52.9 minutes, plus 744 short synthesized confirmation clips — collected from Dashers.
• 48 kHz mono WAV. Per-clip conversational-role annotations included, and transcripts for all 1,293 human-voice recordings. Single test split of 248 ordering tasks across 15 restaurants.
• A further 1,580 recordings / 70.5 minutes of restaurant ambience, played as background noise between conversational turns.
• Language: English. Recordings were collected in the United States.

How it was collected

The recordings were contributed by participants who consented to the collection and to this use. The dataset was assembled for the purpose of benchmarking speech and speech-to-speech models.

Intended use

Evaluating and benchmarking automatic speech recognition and speech-to-speech systems. Not intended for, and licensed against, biometric identification or re-identification of any individual (see the use restriction above).

What this dataset is not

The dataset is provided "as is." It is not represented or warranted to be anonymized or de-identified, nor to be free of the voices or personal information of persons other than consenting participants. See Section 6 of LICENSE for the full disclaimer. The license text controls; this card is a plain-language summary only.

License

LICENSE — DoorDash-Audio-Benchmarks Dataset License, Version 1.0. Do not tag this dataset with a standard open-source identifier; use LicenseRef-DoorDash-Audio-Benchmarks-Dataset-License-1.0.

Citation

If you use this dataset, please cite it as:

DoorDash, Inc. (2026). DoorDash-Audio-Benchmarks. https://huggingface.co/datasets/doordash/doordash-audio-benchmarks

Contact

Questions about the dataset can be filed as issues in the project's GitHub repository.

---

Dataset structure

Each row is one ordering task: the customer's voice clips, the menu, the ground-truth expected_cart to score against, and the conversation decision_tree.

from datasets import load_dataset
ds = load_dataset("doordash/doordash-audio-benchmarks", split="test")

Reading a clip's audio decodes it, which requires torchcodec — without it datasets raises ImportError: To support decoding audio data, please install 'torchcodec'. Install it, or append .decode(False) to get the encoded WAV bytes instead:

ds = load_dataset("doordash/doordash-audio-benchmarks", split="test").decode(False)
row = next(iter(ds))
open("clip.wav", "wb").write(row["clips"][0]["audio"]["bytes"])

A benchmark run plays the customer clips at a speech-to-speech model acting as the restaurant's order taker, then scores the cart the model assembled against expected_cart. The decision_tree selects which clip plays next in response to what the model says, so the customer side is reactive rather than a fixed script.

Field	Type	Description
task_id	string	Task identifier, orderup-0000 … orderup-0253. Not contiguous: 6 ids were retired when their tasks were withdrawn, and surviving ids were not renumbered
restaurant	string	Restaurant name for the task's menu
language	string	Spoken language (English for all current rows)
clips	list	Ordered list of clip structs — see below
expected_cart	JSON string	Ground-truth cart to score model output against
menu	JSON string	Menu available to the model
system_prompt	string	Order-taker system prompt (plain text, not JSON)
tools	JSON string	Tool schemas exposed to the model
decision_tree	JSON string	Conversation graph: nodes, initial node, noise rotation

Clips

audio holds encoded WAV bytes. Each clip struct carries:

Field	Description
clip_id	Filename stem. The decision_tree references clips by <clip_id>.wav
clip_kind	Fine-grained role: main, opt*, restatement*, *_corr, *_post_revision, tts_filler
clip_role	Coarse role: one of main, opt, restatement, corr, yes, no, or null. Null is meaningful — every *_post_revision clip has no role
clip_source	Which family the clip belongs to: manifest, fallback, option, sub_clip, tts_filler, orphan. Also marks this schema version; rows published before it lack the field
transcript	What is spoken. Empty for tts_filler, whose content is its clip_id
item_index, stage_index	Which order item / conversation stage the clip belongs to
audio_duration	Length in seconds, measured from the shipped audio
customer_spans	Source spans: label, text, item_indices, and t_start/t_end
muted_intervals, post_revision_offset_t	Ranges replaced by ambience; offset of the post-revision portion
timing_unreliable	See the timing note below

Timing is clip-relative. audio_duration and every time in customer_spans, muted_intervals and post_revision_offset_t are measured from the start of the clip's own audio, not from the source recording. Absolute positions within the unpublished session recordings are deliberately not included. Durations and the gaps between spans — which is what scoring uses — are unaffected.

A span time may be negative or exceed audio_duration. That is not an error: the clip's audio boundaries were refined after the spans were transcribed, so a span can describe speech that sits partly or wholly outside the clip that references it.

For 7 clips the recorded audio window disagrees with the rendered audio, leaving no reliable origin to measure from. Those carry timing_unreliable: true, and all of their clip-relative times are null. audio_duration remains accurate — it is measured from the audio itself.

Sidecar audio is published as its own clip whose clip_id carries a suffix: <clip_id>.fallback (a shortened re-read), <clip_id>.option (an inline option phrase), and <clip_id>.sub_i<N> (one item's speech cut out of a clip covering several). Writing each clip to <clip_id>.wav reproduces exactly the filenames the decision tree expects.

Every task also carries three tts_filler clips — yes, no, and thats_all — which are synthesized rather than recorded, and are used for confirmations and order close-out.

Decision tree

Each decision_tree contains version, task_id, clips_base_dir, initial_node, initial_noise_rotation, and nodes. Node IDs encode game and stage position (e.g. g1_s0_i0_main). initial_noise_rotation names the ambient background-noise files a run mixes in between clips. A node may also carry fallback_clip and fallback_clips, which name the audio to replay when the order taker asks for a repeat or leaves an item unacknowledged.

Files outside the Parquet rows

Three things a benchmark run needs do not fit the one-row-per-task table, and ship as plain files in this repository.

Path	Contents
background_noise/<session_id>/noise_*.wav	1,580 ambience clips, 70.5 min total, 48 kHz mono, across 64 recording sessions. Shared across the tasks recorded in the same session, which is why they are not per-row.
clips/order_up_shared/*.wav	7 clips the customer side can play for any task: just_the_defaults.wav, no.wav, only_one.wav, thats_all.wav, the_first_option.wav, the_second_option.wav, yes.wav. Not per-row because they are identical everywhere.
task_id_map.json	Per task: session_id, game_index, source_window_seconds, and the original internal task id.

Background noise is required, not decorative. Every decision tree names the ambience files it expects — initial_noise_rotation for the stretch before the first clip, and a per-node noise_rotation thereafter — with the sequence fixed at build time so the noise between any two clips is identical across runs of the same task. A run without these files plays digital silence instead, which changes how a model's voice-activity detection endpoints each turn. Published results were produced with them loaded.

The shared clips are required too. The customer side offers a fixed set of responses — answering yes or no, declining an extra unit, picking the first or second option offered, accepting default options, and closing the order out. Each is one file in clips/order_up_shared/. A missing file does not error; it silently removes that response from the customer's repertoire, so the model is never asked the question it was supposed to answer.

task_id_map.json is what makes the noise locatable: the noise directory is keyed by recording session, and the published task_id values do not encode it. The same file lets a published result be traced back to the row that produced it.

source_window_seconds is the length of the conversation each task was cut from. A harness should scale it to decide how long to let a run go — a task cut from a three-minute exchange needs more than a fixed short timeout, or the conversation is truncated before the order is finished and scored as a failure that never happened.

Limitations

• English only; no other languages in this release.
• 248 tasks across 15 restaurants — menu and restaurant diversity is narrow relative to real ordering traffic.
• Customer turns are prerecorded and selected by the decision tree, so runs are reactive but not fully open-ended conversation.
• yes, no, and thats_all are synthesized, not recorded speech.
• This is a single-domain, task-oriented benchmark. Results on it should not be read as a general measure of voice-agent or speech-recognition quality.

Changelog

This revision

• 6 tasks withdrawn (one recording session removed). The split is now 248 tasks; their task_id values are retired rather than reused, so ids are no longer contiguous.
• Added only_one.wav to clips/order_up_shared/. It was missing from the previous revision, which silently removed the customer's ability to decline an extra unit of an item already ordered.
• Per-clip metadata extended with clip_role, clip_source, item_index, stage_index, audio_duration, customer_spans, muted_intervals, post_revision_offset_t and timing_unreliable. Additive: earlier consumers that read only clip_id/clip_kind/transcript/audio are unaffected.
• Sub-clip replay audio (<clip_id>.sub_i<N>) published for the tasks whose decision trees reference it, so per-item replay resolves instead of falling back to replaying a whole multi-item clip.
• Legacy resp* clip kinds and node ids renamed to opt*. Self-consistent within a row, but any external artifact keyed on an old clip id will not match.