Projects
Out of the lab.
Benchmark ยท September 2026
Logatra Benchmark
Which speech-to-text engine actually hears India? Seven engines on twelve Indic languages and Hinglish, plus fresh 2026 audio no model has trained on. Every raw response is public.
- Engines
- 7
- Languages
- 12
- Fresh clips
- 132
Code, data and method
MIT-licensed harness, manifests, the adjudication log, and the raw provider responses, so the scoring can be checked offline. The Hugging Face dataset is next.
View on GitHubNext up
Where it goes from here
- Human transcription of the held-out split, silver to gold.
- Deepgram, ElevenLabs Scribe, Google and AssemblyAI runs.
- Fresh audio beyond Hindi, and speaker diarization.