mrfakename/parakeet-ultra-ONNX

parakeet-ultra ONNX (browser / WebGPU)

ONNX export of moondream/parakeet-ultra (627M, Parakeet TDT, 25 European languages), prepared for onnxruntime-web on the WebGPU execution provider. Used by the Parakeet Redux WebGPU demo.

Based on the fp32 export in Olicorne/parakeet-tdt-0.6b-v3-ultra-onnx. Changes:

• encoder-model.q4.onnx (393 MB): pointwise 1x1 Convs rewritten as MatMul, then 4-bit MatMulNBits (RTN, block 32, asymmetric). Runs on WebGPU.
• encoder-model.q8.onnx (650 MB): the same, 8-bit (block 64), closest to fp32.
• decoder_joint-model.int8.onnx: dynamic int8 decoder + joint. decoder_joint-model.onnx is fp32.
• preprocessor.onnx: NeMo 128-bin log-mel frontend with the STFT cast to float32 instead of float64 so it runs in onnxruntime-web (max abs diff 2e-5 vs the original).

Quality

Measured on 23 clips (JFK, LibriSpeech test-clean, MLS de/fr/es/it/nl/pt/pl), greedy TDT decoding, CPU onnxruntime:

encoder	decoder	WER en	WER MLS	encoder cosine vs fp32
fp32	fp32	3.56	4.73	1.0
fp32	int8	3.91	4.50	1.0
q8	fp32	3.56	4.95	0.9996
q4 (rtn, b32, asym)	fp32	3.20	4.50	0.977

The test set is small, so differences under ~0.5 WER are noise.

I/O

• preprocessor: waveforms f32 [1,N] (16 kHz mono), waveforms_lens i64 [1] → features [1,128,T], features_lens
• encoder: audio_signal [1,128,T], length i64 → outputs [1,1024,T/8], encoded_lengths
• decoder_joint: encoder_outputs [1,1024,1], targets i32 [1,1], target_length i32 [1], input_states_1/2 [2,1,640] → outputs [1,1,1,8198] (8193 token logits, blank = 8192, then 5 duration logits for durations 0–4), output_states_1/2

License

CC-BY-4.0, same as moondream/parakeet-ultra, which builds on NVIDIA's parakeet-tdt-0.6b-v3.