Parakeet TDT 0.6B v3 ONNX
ONNX export of nvidia/parakeet-tdt-0.6b-v3, the multilingual (25 European languages) FastConformer-TDT checkpoint.
Model source
- Kind
- Model
- File size
- 79.4 KB–842 MB
- Version
- vRevision a912f69
- Updated
- Oct 26, 2026
- Downloads
- 12
- Source
- Hugging Face
Source description
Parakeet-TDT-0.6B-v3 — ONNX
ONNX export of nvidia/parakeet-tdt-0.6b-v3,
the multilingual (25 European languages) FastConformer-TDT checkpoint.
How to tell this apart from the English-only export
vocab.json holds 8192 tokens.
This matters more than it sounds. The English-only Parakeet checkpoint exports to byte-identical file names through the identical code path, and differs only in vocabulary size — 1024 English word pieces against 8192 multilingual ones. Load the English model and feed it German and it does not fail: it forces German phonetics through English word pieces and returns confident nonsense ("Nerval akien Enlich Vivald soda hundred") that looks like a transcript to everything downstream.
So: assert the vocab size at load time. A 1024-token vocab in a bundle that claims to be this one is the English model, whatever the file names say.
| repo | vocab | languages |
|---|---|---|
soniqo/Parakeet-TDT-0.6B-ONNX | 1024 | English only |
soniqo/Parakeet-TDT-0.6B-v3-ONNX (this) | 8192 | 25 European |
Files
| file | precision | notes |
|---|---|---|
parakeet-encoder.onnx | INT8 (dynamic, per-channel) | FastConformer encoder |
parakeet-decoder-joint.onnx (+ .data) | FP32 | fused decoder + joint |
vocab.json | — | 8192 tokens, id → token |
config.json | — | vocab size, blank id, TDT durations, precision |
The FP32 encoder (parakeet-encoder.onnx + .data, 2.5 GB) lives at earlier
revisions of this repo; pin one if you need full precision (e.g. for GPU
execution providers, where integer weight-only graphs give no benefit).
Precision
The encoder is weight-only dynamic INT8: MatMul weights in signed int8 with per-channel scales and reduce_range (7-bit weights — required so the AVX2 u8s8 kernels on pre-VNNI x86 CPUs cannot saturate their int16 accumulators), activations quantized at runtime (no calibration). Attention activation×activation MatMuls and all convolutions stay FP32. Measured against the FP32 encoder on the same audio, references, and scorers (English, German, French, German–English code-switch): mean −0.10 WER, worst slice +0.00 — parity or better on every slice. An earlier per-tensor variant cost +0.1..+0.4 WER; per-channel scales close that gap. Runtime memory roughly halves versus FP32.
The decoder-joint stays FP32 deliberately. Its LSTM hidden state drifts under INT8 and the logit error is large enough to emit blanks where tokens belong, silently dropping whole phrases. Quantize the encoder, never the decoder.
Input
Raw float32 PCM, 16 kHz, mono → 128-bin log-mel → encoder.
Sources
1 source
Sources
1 sourceVerified Oct 26
Download files
Model artifacts
2 artifacts
Model artifacts
2 artifactsSource excerpts
3 excerpts
Source excerpts
3 excerpts--- license: cc-by-4.0 library_name: onnx pipeline_tag: automatic-speech-recognition tags: - onnx - parakeet_tdt - automatic-speech-recognition - speech language: - en - de - fr - es - it - pt - nl - pl - ro - hu - cs - sk - sl - hr - bg - da - et - fi - el - lv - lt - mt - sv - uk - ru --- # Parakeet-TDT-0.6B-**v3** — ONNX ONNX export of [`nvidia/parakeet-tdt-0.6b-v3`](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3), the **multilingual** (25 European languages) FastConformer-TDT checkpoint. ## How to tell this apart from the English-only export `vocab.json` holds **8192** tokens. This matters more than it sounds. The English-only Parakeet checkpoint exports to byte-identical file names through the identical code path, and differs only in vocabulary size — **1024** English word pieces against **8192** multilingual ones. Load the English model and feed it German and it does not fail: it forces German phonetics through English word pieces and returns confident nonsense ("Nerval akien Enlich Vivald soda hundred") that looks like a transcript to everything downstream. So: **assert the vocab size at load time.** A 1024-token vocab in a bundle that claims to be this one is the English model, whatever the file names say. | repo | vocab | languages | |---|---|---| | `soniqo/Parakeet-TDT-0.6B-ONNX` | 1024 | English only | | `soniqo/Parakeet-TDT-0.6B-v3-ONNX` (this) | 8192 | 25 European | ## Files | file | precision | notes | |---|---|---| | `parakeet-encoder.onnx` | INT8 (dynamic, per-channel) | FastConformer encoder | | `parakeet-decoder-joint.onnx` (+ `.data`) | FP32 | fused decoder + joint | | `vocab.json` | — | 8192 tokens, id → token | | `config.json` | — | vocab size, blank id, TDT durations, precision | The FP32 encoder (`parakeet-encoder.onnx` + `.data`, 2.5 GB) lives at earlier revisions of this repo; pin one if you need full precision (e.g. for GPU execution providers, where integer weight-only graphs give no benefit). ## Precision The encoder is weight-only dynamic INT8: MatMul weights in signed int8 with **per-channel** scales and **reduce_range** (7-bit weights — required so the AVX2 u8s8 kernels on pre-VNNI x86 CPUs cannot saturate their int16 accumulators), activations quantized at runtime (no calibration). Attention activation×activation MatMuls and all convolutions stay FP32. Measured against the FP32 e...
Source context: 12 downloads · 0 likes · Pipeline automatic-speech-recognition · Library onnx · Repo soniqo/Parakeet-TDT-0.6B-v3-ONNX
Source context: 58 downloads · 0 likes · Pipeline automatic-speech-recognition · Library onnx · Repo soniqo/Parakeet-TDT-0.6B-v3-ONNX
