| --- |
| language: |
| - hi |
| - bn |
| - gu |
| - kn |
| - ml |
| - mr |
| - or |
| - pa |
| - ta |
| - te |
| - en |
| license: other |
| license_name: proprietary |
| license_link: https://fonadalabs.ai |
| pipeline_tag: automatic-speech-recognition |
| tags: |
| - automatic-speech-recognition |
| - speech-to-text |
| - indic |
| - hindi |
| - api |
| --- |
| |
| # Kapture v3 |
|
|
| Kapture v3 is FonadaLabs' speech recognition model for ten Indian languages and English, built for |
| real conversational audio: call-center recordings, voice notes and voice agents. It is offered as an |
| API; the model weights are not distributed and this repository contains no model files. |
|
|
| - **Docs:** https://fonadalabs.ai/docs/speech-to-text/kapture-v3 |
| - **API key:** https://fonadalabs.ai |
| - **Languages:** Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu, English |
| - **Output:** verbatim transcripts in each language's native script, with segment timestamps |
| - **Audio length:** no limit; long recordings are split and transcribed in parallel |
|
|
| ## Usage |
|
|
| ```bash |
| curl -X POST https://api.fonada.ai/v3/asr/transcribe \ |
| -H "Authorization: Bearer $FONADA_API_KEY" \ |
| -F "file=@call.wav" \ |
| -F "language_id=hi" |
| ``` |
|
|
| ```python |
| import os, requests |
| |
| with open("call.wav", "rb") as audio: |
| r = requests.post( |
| "https://api.fonada.ai/v3/asr/transcribe", |
| headers={"Authorization": f"Bearer {os.environ['FONADA_API_KEY']}"}, |
| files={"file": audio}, |
| data={"language_id": "hi"}, |
| ) |
| print(r.json()["text"]) |
| ``` |
|
|
| ## Results |
|
|
| Self-reported, measured through the production API on 2026-10-22 with each benchmark's official |
| scorer. Word error rate in %, lower is better. |
|
|
| | Benchmark | Metric | Kapture v3 | |
| |---|---|---| |
| | [Voice Arena Monsoon hi-IN](https://huggingface.co/datasets/VoiceArena/Monsoon_hi_test) | OI-WER | 6.21 | |
| | [Vaani Benchmark V1.0](https://huggingface.co/datasets/ARTPARK-IISc/Vaani-Benchmark-V1.0) (Hindi) | WER, 3 references | 13.1 | |
| | [Vistaar](https://github.com/AI4Bharat/vistaar), 8,055-clip sample, 10 languages | WER, Sarvam normalizer | 17.43 | |
|
|
| Open ASR Leaderboard results will be added once the leaderboard maintainers have evaluated the API. |
|
|
| Median server processing time for a short clip is about 0.1 s. |
|
|
