Sahayak Voice Models
The full model bundle behind the Sahayak voice banking assistant: TTS, STT, and
LLM, one folder each. tts/ and stt/ are original fine-tunes; llm/'s base
weights are an unmodified third-party public model (only its Modelfile is original).
All are included here so this repo is a complete, self-contained, reproducible bundle
of everything the assistant runs. Attribution for each is below.
tts/ — original fine-tune
| File | What it is |
|---|---|
t3_nepali_checkpoint.pt |
Fine-tuned T3 (text→speech-token) model, trained on Nepali |
tokenizer_np.json |
Nepali BPE+grapheme tokenizer for the fine-tuned T3 model |
This is a Nepali fine-tune of the T3 stage of
ResembleAI/chatterbox. You still need
the base model's other components to use it — ve.safetensors (voice encoder),
s3gen.safetensors (mel generator), conds.pt (default voice conditioning) — download
those from the original ResembleAI repo; they are unmodified upstream weights and
aren't re-hosted here.
stt/ — fine-tuned on the Hindi version
Based on ai4bharat/indic-conformer-600m-multilingual (ONNX conformer encoder + per-language CTC/RNNT decoder heads, 22 Indic languages including Nepali) — further fine-tuned here on the Hindi version of the model to adapt it for this assistant. All credit for the base architecture/weights belongs to the original AI4Bharat repo; the fine-tuning on top of it was done for this project.
llm/ — base weights are a mirror; the Modelfile is original
| File | What it is |
|---|---|
himalaya-gemma-toolcall.gguf |
Same weights as himalaya-ai/himalaya-gemma-4-e2b-it-gguf — not modified/fine-tuned further here |
Modelfile |
An Ollama Modelfile adding a tool-calling prompt template (JSON function-call format) on top of the base model — this is the only original artifact in this folder |
The underlying weights are unchanged from Himalaya AI Labs' public Gemma-based Nepali model. All credit and license terms belong to that original repo.
Summary of what's actually original here
tts/t3_nepali_checkpoint.pt + tts/tokenizer_np.json (Nepali TTS fine-tune), the
stt/ model (fine-tuned from the Hindi version of IndicConformer), and llm/Modelfile
were produced for this project. The llm/himalaya-gemma-toolcall.gguf weights are an
unmodified copy of a third-party public model, included for a complete, reproducible
bundle — not a claim of authorship over those weights.