striimit/parakeet-tdt-0.6b-v3-webgpu

Parakeet-TDT-0.6b-v3 · ONNX for browser WebGPU

ONNX package of NVIDIA's parakeet-tdt-0.6b-v3 multilingual speech recognition model (25 European languages), prepared for in-browser inference with onnxruntime-web on WebGPU.

Based on the istupakov/parakeet-tdt-0.6b-v3-onnx export. Differences:

• encoder-model.fp16.onnx — encoder converted to float16 (1.24 GB). In browser WebGPU execution the int8 export produces incorrect results and the fp32 export (2.3 GB) is impractical to download; fp16 runs correctly and fast.
• decoder_joint-model.onnx — decoder + joint network, unmodified (fp32, 69 MB).
• nemo128_conv.onnx — 128-bin log-mel feature extraction as a Conv-based ONNX graph (from the onnx-asr project; runs without the STFT operator, which some runtimes lack).
• vocab.txt — vocabulary (8193 tokens, <blk> = 8192).

Notes for implementers:

• The exported encoder's relative-position embeddings cover roughly 35 seconds of audio — split longer audio into chunks before encoding.
• Greedy TDT decoding is implemented in the calling code: the decoder output has 8198 logits per step = 8193 token logits + 5 duration logits.
• Decoder inputs: targets (int32), LSTM states input_states_1/input_states_2 of shape [2, 1, 640].

License: CC-BY-4.0, inherited from the original NVIDIA model. Please attribute NVIDIA (original model) and istupakov (ONNX export).