Problem
The current SALSA pipeline produces corrected text as its final output:
speech → Whisper ASR → LLM correction → text
For real-world Indic voice applications conversational assistants, Krishi Saathi (farmer voice AI), accessibility tools the pipeline needs to return speech, not text. There is currently no TTS module anywhere in this repo.
What's missing
A tts/ module that takes the corrected transcript and synthesizes it back to speech, completing the full round-trip:
speech → ASR → LLM correction → TTS → speech
What I'm proposing
I am building this missing layer using IndicVoice-82M, a TTS model supporting all 22 Indian languages. The addition would include:
tts/synthesize.py takes corrected text, outputs audio via IndicVoice-82M
tts/evaluate.py measures output quality using NORESQA-MOS on synthesized audio
data_preparation/dump_indicvoices.py adds IndicVoices dataset support (currently missing)
This would also enable a new evaluation: does LLM-corrected text synthesize into measurably better speech than raw ASR output? No paper has measured this yet.
Questions for maintainers
Would you accept a PR for this?
Is there a preferred TTS model or evaluation metric you'd like used?
Happy to discuss before opening the PR
Problem
The current SALSA pipeline produces corrected text as its final output:
speech → Whisper ASR → LLM correction → text
For real-world Indic voice applications conversational assistants, Krishi Saathi (farmer voice AI), accessibility tools the pipeline needs to return speech, not text. There is currently no TTS module anywhere in this repo.
What's missing
A tts/ module that takes the corrected transcript and synthesizes it back to speech, completing the full round-trip:
speech → ASR → LLM correction → TTS → speech
What I'm proposing
I am building this missing layer using IndicVoice-82M, a TTS model supporting all 22 Indian languages. The addition would include:
tts/synthesize.py takes corrected text, outputs audio via IndicVoice-82M
tts/evaluate.py measures output quality using NORESQA-MOS on synthesized audio
data_preparation/dump_indicvoices.py adds IndicVoices dataset support (currently missing)
This would also enable a new evaluation: does LLM-corrected text synthesize into measurably better speech than raw ASR output? No paper has measured this yet.
Questions for maintainers
Would you accept a PR for this?
Is there a preferred TTS model or evaluation metric you'd like used?
Happy to discuss before opening the PR