parakeet-tdt-0.6b-v3 — int8 ONNX (encoder / decoder_joint, opset 17)

int8-quantized ONNX build of nvidia/parakeet-tdt-0.6b-v3 (FastConformer encoder + Token-and-Duration Transducer decoder, 25 European languages, automatic language detection and punctuation), re-hosted for AllPrivate — where every model runs in the visitor's browser and nothing is uploaded.

Verified to run fully in the browser with ONNX Runtime Web (wasm execution provider): the 622 MB int8 encoder loads as a single self-contained graph (no external-data files) and produces transcriptions identical to the onnx-asr CPU reference on German and English test clips.

Files

file	role
`encoder-model.int8.onnx`	FastConformer encoder, int8 QDQ, weights embedded
`decoder_joint-model.int8.onnx`	prediction network + joint, int8 QDQ
`nemo128.onnx`	NeMo-exact log-mel preprocessor (128 mel bins, STFT op)
`vocab.txt`	8193-entry vocabulary (`token id` per line, `<blk>` = 8192)
`config.json`	model type / feature size / subsampling metadata
`fetch_and_verify.py`	reproduces this repo from the pinned upstream revision

I/O contract

nemo128.onnx: waveforms [B, N] f32 (16 kHz mono PCM), waveforms_lens [B] i64 → features [B, 128, T] f32, features_lens [B] i64
encoder-model.int8.onnx: audio_signal [B, 128, T] f32, length [B] i64 → outputs [B, 1024, T′] f32, encoded_lengths [B] i64 (subsampling ×8 → one encoder frame per 80 ms)
decoder_joint-model.int8.onnx: encoder_outputs [1, 1024, 1] f32, targets [1, 1] i32, target_length [1] i32, input_states_1/input_states_2 [2, 1, 640] f32 → outputs [1, 1, 1, 8198] f32 (8193 token logits incl. blank, then 5 TDT duration logits for durations 0–4), output_states_1/output_states_2
Greedy TDT decoding: per frame, emit argmax token if not blank (advance LSTM states only on emission), then advance the frame index by the argmax duration (by 1 when the duration head says 0 but the token was blank, or after 10 emissions on one frame)

Provenance

Source model: nvidia/parakeet-tdt-0.6b-v3 (NeMo checkpoint, CC-BY-4.0)
ONNX export + int8 quantization by the onnx-asr project (NeMo model.export(), onnxruntime dynamic quantization), published as istupakov/parakeet-tdt-0.6b-v3-onnx; files taken unmodified from pinned revision 8f23f0c03c8761650bdb5b40aaf3e40d2c15f1ce
sha256 of every file is pinned in fetch_and_verify.py, which re-downloads the upstream revision and checks the hashes
Decode parity validated against onnx-asr 0.x (CPU provider) on German, English and Spanish clips: token-identical output on the wasm execution provider of ONNX Runtime Web 1.26

License

CC-BY-4.0, inherited from nvidia/parakeet-tdt-0.6b-v3 (© NVIDIA). ONNX conversion by Igor Stupakov. This repo only re-packages the already-published ONNX build for browser use.

Downloads last month: 15