Downloads · 30 days
0
VocaHQ/sherpa-onnx-nemo-parakeet-tdt-ctc-110m-en-int8
sherpa-onnx-nemo-parakeet-tdt-ctc-110m-en-int8 is a automatic speech recognition model from VocaHQ. 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.
An int8 build of NVIDIA's parakeet-tdtctc-110m (CTC branch), for sherpa-onnx's OfflineNemoEncDecCtcModelConfig. It is the on-device English model shipped in the VocaPhone keyboard.
Downloads · 30 days
0
Access
Public
Updated Sep 19, 2026
Repo size
132 MB
Likes
0
Public
Click a slice to open those files.
.onnx132 MB · 100%
From the Hugging Face model README
An int8 build of NVIDIA's parakeet-tdt_ctc-110m
(CTC branch), for sherpa-onnx's
OfflineNemoEncDecCtcModelConfig. It is the on-device English model shipped
in the VocaPhone keyboard.
| File | Bytes | SHA-256 |
|---|---|---|
model.int8.onnx | 131,652,171 | 9177a9146cf32ee0cc8152276ef95116f312018d316be37ccf57f7efea81fc1a |
tokens.txt | 9,953 | 450e56bd2f036fe5b6aa821865838cc5aa9d8b0106134ce9a9ba0664abe6cd10 |
3af92f152d32c836acabf38f4c993bc96b80eb2d. tokens.txt is that
revision's file, unchanged.model.onnx with
ONNX Runtime 1.30.0 (QuantType.QUInt8), metadata carried over. See
quantize.py. The published sherpa-onnx int8 repository for this model
contains no model files, which is why this one exists.LibriSpeech test sets, greedy CTC decoding with sherpa-onnx 1.13.8 on macOS arm64, text normalised with the Open ASR Leaderboard's English normaliser:
| Build | Size | test-clean WER | test-other WER |
|---|---|---|---|
| FP32 (upstream export) | 458 MB | 2.93% | — |
| int8 (this repo) | 132 MB | 3.00% | — |
Quantization costs 0.07 WER points on test-clean and decodes about 30% faster on the same CPU. These are desktop measurements, not phone benchmarks.
import sherpa_onnx
recognizer = sherpa_onnx.OfflineRecognizer.from_nemo_ctc(
model="model.int8.onnx", tokens="tokens.txt", num_threads=2,
)
Output is cased and punctuated English.