Training the Mamba architecture on speech and music data using Determined AI
The author wanted to learn the Mamba architecture through hands-on practice by reproducing an open-source speech synthesis script, and encountered repeated dataset failures that made it hard to produce usable model output.
Three of four candidate datasets failed: Schmidt dialogues were too small causing the model to overfit; SpeechTokenizer discards music making the Taylor Swift dataset unusable; and AI Morgan Freeman audio contained excessive pauses that produced mostly empty model outputs.
Using the Alice in Wonderland audiobook dataset with 4 quantizers, the model produced audio that sounds like the input and is not memorized, with 4 quantizers achieving better training results than 8 quantizers.
Frequently asked questions
What did this team achieve with this AI workflow?
Using the Alice in Wonderland audiobook dataset with 4 quantizers, the model produced audio that sounds like the input and is not memorized, with 4 quantizers achieving better training results than 8 quantizers.
What tools did this team use?
Mamba, SpeechTokenizer, Determined AI, ffmpeg, WandB.
What results were reported?
Mamba model parameters: 12M; Training loss with 8 quantizers: never goes below 2; Training loss with 4 quantizers: finally approaches 1; Optimal quantizer count: 4 (source-reported, not independently verified).
What failed first in this deployment?
Three of four candidate datasets failed: Schmidt dialogues were too small causing the model to overfit; SpeechTokenizer discards music making the Taylor Swift dataset unusable; and AI Morgan Freeman audio contained ex…
How is this workflow AI workflow structured?
Download and split audio data → Normalize and tokenize dataset → Train Mamba model → Track experiments with Determined AI → Test generated audio.