A browser ONNX build of ibm-granite/granite-speech-5.0-470m-turboctc (English, French, German, Spanish, Portuguese, Italian and Japanese, about 470M parameters, Apache 2.0 license), made for running in ONNX Runtime Web on WebGPU.
| File | Size | What it is |
|---|---|---|
model.q4.onnx |
284 MB | Raw 16 kHz audio [1, samples] (not normalized; clips under 0.5 s are padded with silence inside the graph) to CTC logits [1, steps, 16384], one step per 80 ms. Feature extraction is part of the graph and stays fp32; the encoder has 4-bit weights (block 128), and the two output layers 8-bit. |
vocab.json |
0.1 MB | Token pieces as a JSON array (index = id). ▁ is a space; <0xNN> pieces are single UTF-8 bytes. Ids below 255 are special tokens. |
Greedy CTC decoding: take the top token per step, merge repeats, drop the blank (id 0).
Word error rate, Whisper English normalizer:
| Version | LibriSpeech test-clean (2,620 clips) | FLEURS English (150 clips) |
|---|---|---|
| Original model (transformers, fp32) | 1.32% | 4.47% |
| This build | 1.43% | 4.84% |
The conversion scripts are in tools/convert-granite-turboctc of the Baa repository.
