Access to DAI-ASR-I18N is granted on request under the DAI-ASR-I18N Data Use Agreement (research and non-commercial evaluation only). The data contains recordings of identifiable people. You agree to the Data Use Agreement: no re-identification, no redistribution, and deletion of specific recordings when a removal request is propagated to you.

Dataset Card for DAI-ASR-I18N

Publication: https://research.withdavid.ai/blog/dai-asr-i18n
GitHub repo: https://github.com/withdavid-ai/dai-asr-i18n

DAI-ASR-I18N is a benchmark from David AI for speech recognition and speaker diarization on everyday two-speaker conversations across 21 languages. In the audio corpus, each speaker is recorded on a separate channel, and each clip includes human-verified, speaker-attributed verbatim transcripts. The conversations are unscripted and capture natural speech patterns, including interjections and backchanneling. To make the benchmark reproducible and support open speech research, we release a frozen public dataset containing about two hours of audio per language.

Dataset Owner: David AI Labs
Languages: 21 (listed below)
License: custom Data Use Agreement (license: other); see Access and licensing
Release snapshot: 2026-09-26
Scoring code and methodology: the David AI's DAI-ASR-I18N repository, released under the MIT License.

Supported tasks and metrics

Automatic speech recognition. The public headline metric is WER for every benchmark language except Chinese, Japanese, Korean, and Thai, which use CER. Reference and hypothesis text are normalized with the same deterministic per-language policy. For speaker-attributed transcription accuracy we report cpWER/cpCER.
Speaker diarization. Diarization uses DER with its miss, false-alarm, and speaker-confusion components, plus speaker-count metrics.

Languages

Arabic (ar), Bengali (bn), Chinese (zh), English (en), French (fr), German (de), Hindi (hi), Indonesian (id), Italian (it), Japanese (ja), Korean (ko), Marathi (mr), Portuguese (pt), Russian (ru), Spanish (es), Tagalog (tl), Tamil (ta), Telugu (te), Thai (th), Turkish (tr), and Vietnamese (vi).

French (fr), Italian (it), and Korean (ko) are not included in the public release at this time, pending completion of our internal data privacy review.

Dataset statistics

Statistic	Public release
Clips	1,035
Total two-channel audio duration	42.1 hours
Speaker channels	2,070 (2 per clip)
Approximate speaker-channel duration	84.2 hours
Languages	21
Distinct speakers	682
Speaker-channel gender (weighted by speaker channel)	54.4% female, 44.3% male, 1.2% undisclosed

Splits

Public evaluation set (released here): about two hours per language across 21 languages.
Private held-out set (not released): retained to reduce training-data contamination and kept private.

Dataset structure

The dataset is delivered as per-clip audio files plus a metadata and transcript table (one row per clip, keyed by sample_id). Each clip provides two per-speaker channels (ch1, ch2), each a single-channel (mono) track, delivered as lossless FLAC under audio/<sample_id>/{ch1,ch2}.flac, at its native source sample rate (16, 24, 32, or 44.1 kHz; 16-bit). A mixed-speaker mono track is not shipped; we provide the code in our repo for mixing. The table below carries the metadata and the transcript.

Metadata and transcript columns:

Metadata: sample_id (a stable namespaced UUID — uuid5 — of the source id), language, locale, duration_type (short ≤6 min / long ≥14 min), duration_s, n_speakers, topic, overlap_type (low <10% / medium 10–20% / high ≥20%), clip_overlap_frac, speaker1_gender, speaker2_gender, speaker1_id, speaker2_id, benchmark_set. speaker1_id and speaker2_id are stable namespaced UUIDs (uuid5, same namespace as sample_id) for the two speakers: a speaker keeps the same id across every clip they appear in, so clips can be grouped by speaker, but the id cannot be linked back to any real identity. No real speaker identity is released.
Transcript: word_alignments, the human-verified, speaker-attributed transcript carrying both the reference text and its word-level timing, as a list of per-segment structs {speaker, start, end, timing_source, words:[{start, end, text}]}. The reference text for WER/CER/cpWER is the sequence of words[].text; the start/end on each word carry the timing for DER.

word_alignments carries the word-level text and timestamps, produced by language-specific forced alignment with a MeetEval character-based pseudo-timing fallback for segments whose forced alignment failed (timing_source: "pseudo"; otherwise the acoustic forced timing is retained).

{
  "sample_id": "e007337e-293e-5296-aaee-3f05b85c5658",
  "language": "en",
  "locale": "en-US",
  "duration_type": "long",
  "duration_s": 120.4,
  "n_speakers": 2,
  "topic": "...",
  "overlap_type": "low",
  "clip_overlap_frac": 0.07,
  "speaker1_gender": "female",
  "speaker2_gender": "male",
  "speaker1_id": "54984bde-f203-5146-ab7f-28a346faa5ec",
  "speaker2_id": "b9b671d4-a2bc-5c3d-8b15-169462340343",
  "benchmark_set": "public",
  "word_alignments": [
    {"speaker": 1, "start": 0.42, "end": 2.81, "timing_source": null,
     "words": [{"start": 0.42, "end": 0.71, "text": "..."}]},
    {"speaker": 2, "start": 2.95, "end": 4.10, "timing_source": "pseudo",
     "words": [{"start": 2.95, "end": 3.30, "text": "..."}]}
  ]
}

Dataset creation

Source data. Natural two-party conversations recorded per speaker channel by contributors engaged for data collection.
Transcripts. Human-verified verbatim transcripts, speaker-attributed at the segment level.
Word-level timestamps. We used language-specific forced-alignment algorithms to obtain word-level timestamps for the transcript in order to evaluate time-based metrics (DER). Segments whose forced alignment failed fall back to MeetEval character-based pseudo timing, recorded as timing_source: "pseudo".

Data access and privacy

Recipients must follow the approved Data Use Agreement and all privacy laws applicable to their processing.

Do not attempt to identify, locate, or contact a speaker. Report suspected sensitive-content or consent issues to research@withdavid.ai. The release's consent-withdrawal and erasure procedure, including recipient-notification obligations, is defined in the Data Use Agreement.

Important notes:

Evaluation only, non-commercial. This is a test benchmark, not training data.
Coverage and balance. Balanced to about two hours per language and roughly balanced by gender.
Data collection. The data collection process follows a rigorous standard under applicable privacy laws; individual contributors are compensated at a fair market rate in their local regions.

Access and licensing

Data: access is gated and granted under the DAI-ASR-I18N Data Use Agreement, version 1.0, dated 2026-09-30 (license: other). The agreement limits use to approved non-commercial research and evaluation, prohibits re-identification and redistribution, and defines recipient obligations for propagated removal requests. Requesting Hub access does not replace the agreement; access is granted only after the required agreement and review steps are complete.
Code: the evaluation harness and scoring code are released separately under the MIT License in the David AI ASR Bench repository (https://github.com/withdavid-ai/dai-asr-i18n).

Citation

@techreport{davidai2026dai-asr-i18n,
  title       = {DAI-ASR-I18N: A Multilingual, Dual-Channel Conversational Benchmark for Speech Recognition and Speaker Diarization},
  author      = {{David AI}},
  institution = {David AI},
  year        = {2026},
  url         = {https://research.withdavid.ai/},
}

Downloads last month: 103

Total file size:

4.09 GB