Training for other languages
Do you have a tutorial to support other languages?
Working on it!
Happy to help with this — I'd like to bring NeuDecide to Russian.
From reading the Space demo, the language-dependent part looks like the audio encoder only: tool schemas and the output JSON stay language-neutral, so a new language shouldn't need tokenizer or text-side changes. My plan for a pilot:
Generate synthetic training pairs — Russian command utterances (via an open Russian TTS) paired with tool calls in the same JSON-Schema format used for English.
Adapt the audio encoder (+ decoder) on those.
Evaluate on real Russian speech, not the TTS used for training, against gold tool calls.
A few questions:
Any plan to release the training recipe? The Space README references an export/ pipeline, but it's not in the public repo — architecture is recoverable from the ONNX graphs, the data recipe and hyperparameters are the hard part.
What is the audio encoder pretrained on — does it include multilingual speech data?
Roughly how much synthetic data went into the English model, so I can size the effort?
Is per-language fine-tuning from your base checkpoint the intended path, or a full retrain?
Happy to share results and a write-up if the pilot works.