Facing accuracy issues while transcribing Hindi (indian regional language)

[nagarajuprimefocus]

Currently accuracy for Hindi is not good. Any suggestion to move in right direction to achieve good accuracy ?

[SandraRodgers]

Hi @nagarajuprimefocus ,

I recommend you try changing the tier you are using. Hindi is available on the base or enhanced tier.

https://api.deepgram.com/v1/listen?tier=enhanced&language=hi

The other option would be to try Deepgram's Whisper Cloud offering:

https://api.deepgram.com/v1/listen?model=whisper&detect_language=true

or try:

https://api.deepgram.com/v1/listen?model=whisper&language=hi

[rahulbansal16]

Hi @SandraRodgers
Is it possible to setup deepgram for this workflow?

The input audio can be in either Hindi or English.
If the input is in English, output the English transcript,
but if the output is Hindi, output the Hindi in Latin transcript.

I don't want to manually pass the language for each audio file.

[jkroll-deepgram]

Hi @rahulbansal16, with both Deepgram's enhanced and base model tiers and with our managed Whisper offering, you can use detect_language=true, and the model will detect whether each file is in Hindi or English, and transcribe it in that language.

https://api.deepgram.com/v1/listen?tier=enhanced&detect_language=true
https://api.deepgram.com/v1/listen?model=whisper&detect_language=true

Unfortunately however, the current default for Hindi is in Devanagari script, so if Hindi is detected, it will transcribe in Devanagari (hi language code). The only way to get Latin alphabet transcription is to know the file is in Hindi, and specify language=hi-Latn.

We understand that this is not ideal, and are exploring a future feature that will give customers more control over the languages and scripts that can be detected.