mistralai

/

Voxtral-4B-TTS-2603

like 759

Follow

Mistral AI_ 16.9k

Environment

1. Model: voxtral-mini-tts-2603 (via Mistral API)
2. API endpoint: /v1/audio/speech
3. Response format: mp3
4. Use case: Multi-speaker podcast generation (open-notebooklm; 3 speakers, 38 dialogue clips concatenated)

Issues
I'm using Voxtral TTS to generate a multi-speaker podcast by synthesizing each dialogue turn individually (one API call per turn) and then concatenating the clips. I've encountered three issues:

1. Speech truncation at the end of utterances
   Some speakers' speech is cut off before they finish their sentence. The final words or syllables are abruptly truncated, making the dialogue sound incomplete. This happens inconsistently — not every clip is affected, but it's noticeable enough to degrade the listening experience.

2. Unnatural silence/padding at clip boundaries
   Each generated audio clip contains noticeable silence at the beginning and/or end. When clips are concatenated sequentially, these silences accumulate and create unnaturally long pauses between speaker turns. The gaps feel much longer than what you'd hear in a real conversation.

3. Volume fluctuation across different speakers/voice_ids
   When switching between different voice_ids (different speakers), the output volume varies significantly. Some clips are noticeably louder or quieter than others, even though all other parameters remain the same. This creates jarring volume jumps in the final podcast.