Fonada Labs
Indic speech benchmark
Fonada Labs publishes these operating figures for its Indic speech models. They describe that model set. They are not a leaderboard against other speech companies.
The models are on fonadalabs.ai. Fonada AI voice agents that use Indian-language speech are on the voice channel.
Languages
23 Indian languages
Speech recognition coverage published by Fonada Labs. Streaming recognition.
Clean audio
Above 95%
Recognition on quiet, clean recordings. This is easier than a phone call.
Telephony audio
Above 90%
Recognition on phone-line audio. Phone audio is harder than a clean recording.
Speech-path latency
Under 200 ms
Fonada Labs figure for the speech path. It is not a ranking against another vendor.
Text-to-speech catalog
Fonada Labs lists 111 voices across Hindi, Tamil, Telugu, and English, delivered as 24 kHz MP3. That catalog is the Labs text-to-speech set, separate from the recognition figures above.
How to read the two accuracy figures
Clean audio is a quiet recording. Telephony audio is a phone call, with the noise and compression of a phone line. A model that is above 95% on clean audio can be lower on a call. Fonada Labs publishes the telephony figure on its own: above 90%.
Use the telephony figure when the audio is a phone call. Use the clean-audio figure for a studio or headset recording. Do not treat either number as a score versus another vendor.
What Fonada AI voice agents support
On fonada.ai, a voice agent supports 60+ languages, including 35+ Indian languages, for recognition and speech. That is the agent coverage. The 23-language, 111-voice, and accuracy figures on this page are the Fonada Labs model figures. Hosting for the Labs speech stack is in India, including Bengaluru, Mumbai, and Noida.
Questions
What does this Indic speech benchmark measure?
It reports Fonada Labs operating figures for Indic speech recognition and text-to-speech: 23 Indian languages, recognition above 95% on clean audio and above 90% on telephony audio, 111 voices, and speech-path latency under 200 ms. It does not score other companies.
Is this a comparison with Whisper, Google, or another speech model?
No. These are Fonada Labs figures for its own models. This page does not publish word-error rates for any other vendor.
Where are the Fonada Labs models?
The speech recognition and text-to-speech models are on fonadalabs.ai. Fonada AI voice agents, on fonada.ai, support 60+ languages, including 35+ Indian languages.
Why are clean-audio and telephony figures different?
A quiet recording is easier to recognize than a phone call. Fonada Labs publishes both: above 95% on clean audio and above 90% on telephony audio.