kitten-asr-tiny (ONNX export)
ONNX export of KittenML/kitten-asr-tiny,
built and consumed by itamaker/kitten-asr-go,
a Go speech-to-text engine. KittenML's original repo ships only PyTorch
safetensors weights, no ONNX; these files are produced by that project's
tools/ export pipeline.
This is not an official KittenML release — it's an independent, unofficial conversion. A dynamically-quantized int8 export of this same model is also available — about 3.8-3.9x smaller on disk, at a measured accuracy cost (see that repo's card for the tradeoff).
Files
| File | Purpose |
|---|---|
kitten_asr_tiny_audio_encoder.onnx | Fixed 30s/3000-mel-frame audio encoder graph |
kitten_asr_tiny_decoder.onnx + decoder.onnx.data | Autoregressive text decoder with explicit per-layer KV-cache tensors |
kitten_asr_tiny_embed_tokens.bin | Flat float32 embedding table, loaded directly (not an ONNX graph) |
config.json, vocab.json, merges.txt, added_tokens.json | Tokenizer/model config, copied from the original checkpoint |
Usage
git clone https://github.com/itamaker/kitten-asr-go
cd kitten-asr-go
HF_TOKEN=hf_xxx scripts/fetch_model.sh tiny
go build -o bin/ ./...
./bin/kitten-asr models/kitten-asr-tiny-onnx hello.wav
See itamaker/kitten-asr-go for the CLI, HTTP server, and Go library docs.
License
KittenML has not published license terms for kitten-asr-tiny's weights as
of this writing (no license field, no model card on the original repo).
This repo makes no license claim of its own for the weights — check with
KittenML directly before relying on them for anything beyond evaluation.