[1/4] How well do speech recognition systems handle real conversations in 21 languages? Today we're releasing DAI-ASR-I18N, a benchmark that evaluates 14 systems on speech recognition and speaker diarization in natural, unscripted conversation

Oct 1, 2026 · 2:48 AM UTC

6
4
5
37
19,921
RelevantRecentLikes
[2/4] What we found - part 1 1. English scores hide big gaps between systems in other languages 2. Microsoft AI’s MAI-Transcribe-2 leads ASR in 18 of 21 languages; ElevenLabs' Scribe v2 leads in the other three
1
4
438
[3/4] What we found - part 2 3. Every dedicated diarization model beat every combined transcription-and-diarization system; NVIDIA Nemotron 3 Diarization leads in 19 of 21 languages 4. Every system struggles most with overlapping speech
1
1
4
372
[4/4] We're releasing the dataset and our inference and evaluation harness publicly. Results: research.withdavid.ai/blog/a… Data: huggingface.co/datasets/davi… Evals or research collaboration: [email protected] [email protected]

DAI-ASR-I18N: Benchmarking Multilingual Conversational ASR

research.withdavid.ai
3
274
We love seeing Nemotron 3 Diarization #1 😍
14
934
This is great! I'd love to see the same analysis for the realtime/live variants of these systems. Do you plan to do that too?
7
A 21-language benchmark on unscripted conversation is the right test, since most published WER comes from read speech. For anyone choosing a system: read the diarization error separately from the word error. They diverge hard on crosstalk. brss.fyi/1ux8

AI Transcription with Speaker ID | BrassTranscripts

AI transcription with speaker ID in 99+ languages. $2.50 for 15 minutes. Professional-grade quality. Start transcribing in minutes.

brasstranscripts.com
51
Wow, great name selection for tts/stt 21:79 from the Holy Quran We guided ˹young˺ Solomon to a fairer settlement, and granted each of them wisdom and knowledge. We subjected the mountains as well as the birds to hymn ˹Our praises˺ along with David. It is We Who did ˹it all˺.
10
Moi, en tout cas, pour moi qui parle français, c’est un banger. Maintenant, moi, ce que j’aimerais savoir, c’est est-ce que tu as pensé à ajouter le nouveau modèle chinois FunASR ? Parce que ce modèle d’Alibaba, il est premier actuellement sur Artificial Analysis Benchmarks.
7