sherpa-onnx Core

1.12.40 verified Mon Apr 27 auth: no python

Core shared libraries for sherpa-onnx, a cross-platform speech recognition and text-to-speech toolkit. Version 1.12.40 provides the underlying ONNX Runtime-based engines for ASR, VAD, and TTS, with package updates for various platforms. Release cadence is approximately semi-monthly.

pip install sherpa-onnx-core

Common errors

error ImportError: cannot import name 'OfflineRecognizer' from 'sherpa_onnx' ↓

cause The 'sherpa-onnx-core' package does not include the Python bindings; only the C libraries.

fix

Install 'sherpa-onnx' (pip install sherpa-onnx) which provides the Python API.

error ModuleNotFoundError: No module named 'sherpa_onnx' ↓

cause Neither 'sherpa-onnx' nor 'sherpa-onnx-core' is installed.

fix

Run 'pip install sherpa-onnx'.

error RuntimeError: Model file not found: /path/to/encoder.onnx ↓

cause The model path is incorrect or the model was not downloaded.

fix

Download the model files from the official model zoo and provide absolute paths.

Warnings

gotcha The 'sherpa-onnx-core' package only contains shared libraries; the Python API is provided by 'sherpa-onnx'. Installing 'sherpa-onnx-core' alone will not give you importable Python modules. ↓

fix Install 'sherpa-onnx' instead: pip install sherpa-onnx

breaking Model file formats may change between major releases. Models downloaded for version 1.9 may not work with 1.12. ↓

fix Always download the latest model files for your sherpa-onnx version.

deprecated The 'sherpa_onnx.OfflineRecognizer' constructor signature changed in v1.10; 'from_models' is now the preferred factory method. ↓

fix Use 'OfflineRecognizer.from_models(...)' instead of 'OfflineRecognizer(...)'.

Imports

OfflineRecognizer
wrong
```
from sherpa_onnx import OfflineRecognizer
```
correct
```
import sherpa_onnx
```
The recommended import is the module-level import; 'from sherpa_onnx import OfflineRecognizer' works but is less conventional.
OnlineRecognizer
wrong
```
from sherpa_onnx_core import OnlineRecognizer
```
correct
```
import sherpa_onnx
```
sherpa-onnx-core does not expose a separate package; everything is under sherpa_onnx.

Quickstart

Basic offline ASR using pre-downloaded ONNX models.

import sherpa_onnx

recognizer = sherpa_onnx.OfflineRecognizer.from_models(
    encoder="path/to/encoder.onnx",
    decoder="path/to/decoder.onnx",
    tokens="path/to/tokens.txt",
)
wave = sherpa_onnx.read_wave("test.wav")
result = recognizer.decode(wave)
print(result.text)