Downloads · 30 days
0
csukuangfj/sherpa-onnx-nemo-parakeet-tdt-0.6b-v2
sherpa-onnx-nemo-parakeet-tdt-0.6b-v2 is a automatic speech recognition model from csukuangfj. Use it when you need speech turned into text. It is set up for sherpa-onnx. The card lists the license as cc-by-4.0.
This package provides NVIDIA's parakeet-tdt-0.6b-v2 for sherpa-onnx. NVIDIA trained the weights; the sherpa-onnx project exported them from the original .nemo checkpoint. It is one offline transducer — encoder, decode…
Downloads · 30 days
0
Access
Public
Updated Nov 7, 2026
Repo size
3.3 GB
Likes
0
Public
Click a slice to open those files.
.weights2.4 GB · 97%
From the Hugging Face model README
This package provides NVIDIA's parakeet-tdt-0.6b-v2 for sherpa-onnx. NVIDIA trained the weights; the sherpa-onnx project exported them from the original .nemo checkpoint. It is one offline transducer — encoder, decoder and joiner — for 16 kHz mono English audio, published as sherpa-onnx-nemo-parakeet-tdt-0.6b-v2 (fp32) and sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8 (int8).
The fp32 package sherpa-onnx-nemo-parakeet-tdt-0.6b-v2/ contains:
encoder.onnx, with its external weights in encoder.weightsdecoder.onnxjoiner.onnxtokens.txttest_wavs/The int8 package sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8/ contains the quantised graphs instead:
encoder.int8.onnxdecoder.int8.onnxjoiner.int8.onnxtokens.txttest_wavs/Input must be 16 kHz mono audio. Set d to the package directory you downloaded:
import sherpa_onnx
import soundfile as sf
d = "sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8" # or the fp32 package; drop .int8 below
recognizer = sherpa_onnx.OfflineRecognizer.from_transducer(
encoder=f"{d}/encoder.int8.onnx",
decoder=f"{d}/decoder.int8.onnx",
joiner=f"{d}/joiner.int8.onnx",
tokens=f"{d}/tokens.txt",
model_type="nemo_transducer",
)
audio, sample_rate = sf.read("audio.wav", dtype="float32", always_2d=True)
assert sample_rate == 16000, sample_rate
stream = recognizer.create_stream()
stream.accept_waveform(16000, audio[:, 0])
recognizer.decode_stream(stream)
print(stream.result.text)
The encoder metadata written by the exporter contains tdt in its url field, so sherpa-onnx loads the joiner as a TDT joiner rather than a plain transducer joiner.
FLEURS test split, 300 deterministic clips per language, punctuation and case stripped, decoded with one harness for all three models; word error rate (WER) in %. These are this project's own numbers, not the model author's, so they are not directly comparable to published benchmarks.
| Language | v2 (this package) |
|---|---|
| English | 6.48 |
v2 is English only. For comparison, v3 measures 11.16 and Ultra 8.51 on the same English clips, so v2 remains the strongest of the three in English. On the languages v2 does not cover, v3 measures 6.84 German, 8.69 French and 4.55 Spanish, and Ultra measures 5.20, 7.48 and 3.81, so Ultra is the strongest of the three in German, French and Spanish.
Short English clips of 15 s or less (282 clips): v2 6.42, v3 11.01, Ultra 8.56.
The weights are NVIDIA's parakeet-tdt-0.6b-v2, trained by NVIDIA and licensed under CC BY 4.0. This ONNX export was produced by the sherpa-onnx project, whose code is licensed under Apache-2.0. NVIDIA does not endorse this package.
The export script is scripts/nemo/parakeet-tdt-0.6b-v2/. The package is built and published by .github/workflows/export-nemo-parakeet-tdt-0.6b-v2.yaml.