EchoLens contains 20,008 recordings of identifiable human speech from 515 consenting participants, released under Stanford IRB protocol #77233. Access is granted automatically to signed-in Hugging Face users who accept the use policy below.

EchoLens use policy

EchoLens contains identifiable human speech. A voice is biometric, and a
speaker with a known reference sample could in principle be re-identified.
Participants were never asked to state names or other direct identifiers. All
515 participants gave informed
consent to public release of their recordings through public research
repositories (Stanford IRB protocol #77233). Your access is conditioned on the
commitments below.

You may use EchoLens for non-commercial research on the fairness,
robustness, and evaluation of speech and audio-language systems, including
training research models, consistent with CC BY-NC 4.0.

You may not

attempt to identify, contact, or infer the legal identity of any speaker;
use the recordings for speaker recognition, speaker verification, voice
biometrics, voice cloning, or synthetic voice generation;
use the recordings in production or deployed voice systems;
treat any recording or demographic label as evidence of a speaker's actual
identity, race, gender, or national origin;
re-host the audio in any ungated venue;
use EchoLens for any commercial purpose without prior written permission
from the authors.

Demographic labels are self-reported and support a bounded Black/White ×
Female/Male audit design. They are not ground truth about any individual.

Participants retain the right to withdraw. On a withdrawal request the
maintainers will remove the affected recordings and issue a new version; you
agree to delete withdrawn recordings from your copies on notification.

Commercial licensing and all other enquiries: alexsdl@law.stanford.edu

EchoLens

A Human-Speech Dataset for Auditing Demographic Sensitivity in Audio-Language Models

📄 Paper (EMNLP 2026 Findings) · 💻 Code

Voice interfaces are increasingly moving away from transcription pipelines toward end-to-end systems that directly respond to audio inputs. This development in turn requires a shift in evaluation methodology away from transcription accuracy and towards more substantive markers such as response validity. We introduce EchoLens, a demographically stratified dataset of 20,008 recordings from 515 adult participants residing in the U.S., with self-reported demographic information across race, gender, age, accent and primary language, among others. Each participant speaks out loud and verbatim a randomized subset of 69 advice-seeking and estimation prompts spanning 11 domains grounded in the American Time Use Survey. Most prompts elicit quantitative responses from the model, allowing for direct analysis of output distributions across demographic subgroups without requiring reliance on LLM-as-a-judge. EchoLens also contains a documented audit protocol, which—in an illustrative application—we use to evaluate six current audio-language models. Under this assessment, we find that racial disparities are more frequent and larger in magnitude than gender disparities, with particular concentration in two models.

The dataset ships with the model outputs from the paper's audit: 386,208 extracted responses from six ALMs under four input conditions.

At a glance


Speakers	515 (249 `new_recruits`, 266 `reconsent`)
Recordings	20,008 wav, 16 kHz mono 16-bit PCM
Audio	6.0 GB
Instrument	74 utterance roles: 69 scenario prompts (each speaker reads a random 35), 2 verbal suffixes, 1 demographic probe, 1 dialect probe, 1 mic check
Race (self-report)	Black 254 · White 257 · Other/Unknown 4
Gender (self-report)	Female 253 · Male 257 · Non-binary 5
Accent	non-SAE 361 · SAE 154
Model outputs	386,208 responses · 6 models × 4 conditions
ASR transcripts	whisper-1 (external) and Gemini (self-transcript)
License	CC BY-NC 4.0

Quickstart

from datasets import load_dataset

ds = load_dataset("alexsdl/EchoLens", split="full")   # config "clips" is the default
ex = ds[0]
ex["audio"]["array"]            # np.float32, 16 kHz mono
ex["audio"]["sampling_rate"]    # 16000
ex["race_simplified"], ex["gender_simplified"], ex["prompt_text"]

The subset the paper's primary analyses use:

strict = ds.filter(lambda r: r["strict_audit_eligible"])   # 500 speakers

Stream instead of downloading 6 GB:

ds = load_dataset("alexsdl/EchoLens", split="full", streaming=True)

Side tables — three further configs:

participants = load_dataset("alexsdl/EchoLens", "participants",    split="full")
responses    = load_dataset("alexsdl/EchoLens", "model_responses", split="full")
wer          = load_dataset("alexsdl/EchoLens", "wer_per_clip",    split="full")

prompts.csv, recordings.csv and splits.csv are not exposed as configs because every column they carry is already joined into clips. They ship as plain CSVs for anyone who wants the un-joined tables:

import pandas as pd
from huggingface_hub import hf_hub_download
prompts = pd.read_csv(hf_hub_download("alexsdl/EchoLens", "data/metadata/prompts.csv",
                                      repo_type="dataset"))

One participant's raw wav files, without the full download:

hf download alexsdl/EchoLens --repo-type dataset \
  --include 'data/audio/new_recruits/P0001/*' --local-dir ./echolens

Version note. The Audio feature changed in datasets 4.0 to torchcodec-backed decoding, so on 4.0+ you also need pip install torchcodec before ex["audio"]["array"] will work. These examples were tested on datasets 2.19 and 5.0.

Repository layout

The layout mirrors the code repository, so downloaded files drop straight into a clone.

data/
├── metadata/            participants.{csv,parquet}, prompts.csv, recordings.csv, splits.csv
├── clips/               clips-*.parquet — packed audio + metadata (the `clips` config)
├── audio/
│   ├── metadata.jsonl   audiofolder manifest, 20,008 rows
│   ├── new_recruits/P####/*.wav
│   └── reconsent/P####/*.wav
└── model_outputs/
    ├── extracted_responses.parquet   canonical normalized responses (386,208 rows)
    ├── responses/<provider>/<model>/<condition>/part-*.jsonl
    ├── transcripts/{external/whisper-1, self/gemini/...}/part-*.jsonl
    ├── disparity/       SMD outputs: CIs, prompt-level, validity rates, BH correction
    └── wer/             per-clip and by-group WER (.csv and .parquet)

Both the parquet shards and the loose wav files are provided. Parquet powers the viewer, load_dataset(), and streaming; the loose files give per-participant access and a bit-exact archival form. They contain identical audio.

To load the loose files locally after downloading:

ds = load_dataset("audiofolder", data_dir="./echolens/data/audio", split="train")

Subsets

splits.csv carries four nested boolean masks:

Mask	Speakers	Meaning
`in_full`	515	every participant
`in_audit_eligible`	506	fits the Black/White × Female/Male design
`in_strict_audit`	500	recommended — adds agreement between the Prolific recruitment cell and self-report
`in_balanced`	496	124 per race × gender cell

Primary analyses in the paper use in_strict_audit.

Columns

The clips config carries one row per recording: clip identifiers (clip_id, participant_id, cohort, question_id, prompt_type), the prompt (prompt_text, scenario, is_quantitative, direction, suffix fields), the three audit axes (race_simplified, gender_simplified, accent_simplified), speaker demographics (age_bin, income_bin, education, english_proficiency, primary_language, residency_state, state_childhood, recording_device, …), and the split masks.

Two things to watch:

question_id is null for all 515 mic_check rows — the mic check is not a prompt.
participant_total_duration_s is per participant, not per clip. It is that speaker's total recording time across all their clips.

The participants config is the full 515-row table including verbatim free-text self-reports. recordings is the 20,008-row clip manifest.

The audit

386,208 extracted responses, six models × four conditions:

Model	Provider
`gemini-3.1-flash-lite-preview`, `gemini-3.1-pro-preview`, `gemini-3.5-flash`	Google
`gpt-audio-1.5`	OpenAI
`Qwen/Qwen2.5-Omni-7B`	Alibaba (local)
`moonshotai/Kimi-Audio-7B-Instruct`	Moonshot (local)

Condition	Model input
`canonical_text_response`	the prompt as text — control
`direct_audio_response`	the participant's audio
`external_transcript_response`	a Whisper transcript of that audio
`self_transcript_response`	the model's own transcript of that audio

extracted_responses.parquet is the flat, normalized table the analysis runs on. The raw per-call JSONL under data/model_outputs/responses/ is provided for provenance but is not exposed as a loadable config: response_raw holds provider-specific nested shapes that Arrow cannot reconcile into one schema.

In those raw rows, audio_path_with_suffix refers to the assembled-audio cache — a scenario clip concatenated with that speaker's own verbal-suffix recordings. That 7.3 GB cache is derived and not distributed; rebuild it with scripts/build_assembled_audio.py in the code repository.

Collection

Participants were recruited on Prolific across four (race × gender) cells, screened to US residents aged 18+, and recorded through a Qualtrics survey with an embedded Voiceform widget. Qualtrics performed the per-participant randomization of 35 scenario prompts out of 69. All audio was standardized to 16 kHz mono 16-bit PCM.

Two cohorts, same instrument:

new_recruits (249) — recorded 2026-04-28/29, consented to public release at the time of recording.
reconsent (266) — recorded 2025-10-13 in the original study, then re-contacted in 2026 for affirmative redistribution consent. Only those who approved are included. Their audio was originally 48 kHz and was downsampled to 16 kHz.

Ethics, consent, and IRB

EchoLens contains identifiable human speech and participant-linked demographic metadata. The study was reviewed under Stanford IRB Protocol #77233, and the released dataset includes only recordings from participants who consented to public redistribution. Participants were informed that their recordings would be processed by AI models, that demographic and device information would be collected for research purposes, and that recordings authorized for release could be shared through public research repositories. They were compensated for participation, and compensation was not tied to the content of their recordings or to model outputs.

The main residual risk is identifiability. Voice recordings are inherently identifying even though speakers were never asked to state names or other direct identifiers, and once recordings are public the research team cannot fully control downstream access, redistribution, or reuse. To reduce ancillary identifiability the release uses opaque participant identifiers, excludes direct platform identifiers and private linkage files, and retains non-released data under secure access restrictions.

EchoLens is intended for research auditing and evaluation of audio-language models. It is not for speaker identification, voice biometrics, production voice-system deployment, or treating recordings as evidence of speaker identity. Those uses fall outside participant consent and are prohibited regardless of license.

Limitations

Demographic labels are self-reported; speech cues are not ground-truth evidence of identity.
The primary audit design is Black/White × Female/Male and is not representative of all speakers or all forms of demographic harm. Five non-binary participants are retained in in_full but fall outside the audit splits.
Participants who self-report Hispanic/Latino are coded as White in race_simplified, and multi-race Black-and-White participants are coded as Black. Both are documented design choices, not the only defensible ones.
Prompts are scripted, so the corpus measures response variation under controlled input, not naturalistic conversational speech.
Accent (non-SAE / SAE) is a descriptive third axis that the pipeline computes but the paper deliberately keeps out of its headline claim.

License

CC BY-NC 4.0. Commercial use — including revenue-generating products, commercial model training, and paid services — requires prior written permission from the authors: alexsdl@law.stanford.edu. See LICENSE.

Citation

The ACL Anthology entry is not yet published. Until then:

@inproceedings{salinas2026echolens,
  title     = {{EchoLens}: A Human-Speech Dataset for Auditing Demographic
               Sensitivity in Audio-Language Models},
  author    = {Salinas, Alejandro and Koenecke, Allison and Nyarko, Julian},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year      = {2026},
  publisher = {Association for Computational Linguistics},
  note      = {To appear}
}

Downloads last month: 179

Total file size:

12.9 GB