Repeatable MMS-TTS comparisons: what should a run record include?
I’m preparing MMS-TTS experiments in Sinaps Studio, our Integrated Model Environment (IME), and looking at how long-text handling should be recorded alongside the worker code.
I found a community report from April 2024 where an extremely large input triggered a token-length warning (510,649 tokens versus a configured limit of 4,096). I haven’t reproduced that case in Studio.
For this checkpoint, what input-length checks and sentence/paragraph chunking policy would you recommend? I’d like to distinguish a tokenizer configuration warning from an actual model or memory constraint before choosing a limit.
For repeatable comparisons, my planned run record includes:
- Model/tokenizer revisions, exact text, normalization and chunk boundaries.
- Seed,
speaking_rate,noise_scaleandnoise_scale_duration, following the VITS documentation. - Transformers/PyTorch versions, device and dtype.
- Output sampling rate and any resampling applied.
Would you reset the seed for each chunk when comparing settings, then repeat across several seeds to check variation?
The screenshot shows the setup/worker editor; it is not an error capture or an audio-quality benchmark.
