parakeet-redux ONNX (browser / WebGPU)

ONNX files for moondream/parakeet-redux (149M, Parakeet TDT, 25 European languages), used by the Parakeet Redux WebGPU demo.

This is a mirror of eschmidbauer/parakeet-redux-onnx, plus one added file:

  • decoder_joint-model.int8.onnx (18 MB): dynamic int8 quantization of the decoder + joint network. It gives the same WER as fp32 on our test clips (en 4.63, MLS 6.08) and decodes about 4Γ— faster on CPU.

The encoder (encoder-model.onnx) runs on onnxruntime-web's WebGPU execution provider. The preprocessor and decoder run on WASM.

I/O

  • preprocessor: waveforms f32 [1,N] (16 kHz mono), waveforms_lens i64 [1] β†’ features [1,128,T], features_lens
  • encoder: audio_signal [1,128,T], length i64 β†’ 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] β†’ outputs [1,1,1,8198] (8193 token logits, blank = 8192, then 5 duration logits), output_states_1/2

License

CC-BY-4.0, same as moondream/parakeet-redux. Original ONNX export by eschmidbauer.

Downloads last month
43
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for mrfakename/parakeet-redux-ONNX

Quantized
(12)
this model

Spaces using mrfakename/parakeet-redux-ONNX 2