Hinotsuba/sherpa-onnx-whisper-tiny

Sherpa-ONNX Whisper Tiny

This repository contains the multilingual OpenAI Whisper Tiny model converted to ONNX format for sherpa-onnx.

It includes both full-precision and INT8-quantized encoder/decoder models for local Speech-to-Text inference.

Repository Contents

• —tiny-encoder.onnx — Whisper Tiny encoder.
• —tiny-decoder.onnx — Whisper Tiny decoder.
• —tiny-encoder.int8.onnx — INT8-quantized encoder.
• —tiny-decoder.int8.onnx — INT8-quantized decoder.
• —tiny-tokens.txt — Whisper tokenizer vocabulary.
• —test_wavs/ — Example audio files for testing.

Model

• —Architecture: OpenAI Whisper Tiny
• —Variant: Multilingual
• —Format: ONNX
• —Runtime: sherpa-onnx / ONNX Runtime
• —Task: Automatic Speech Recognition and speech translation
• —Parameters: ~39M

The INT8 models provide a smaller and potentially faster alternative to the full-precision ONNX models, depending on the target hardware and runtime.

Upstream Projects

The original Whisper architecture and model weights were developed and released by OpenAI:

OpenAI Whisper: https://github.com/openai/whisper

The ONNX export and sherpa-onnx integration originate from the sherpa-onnx project:

sherpa-onnx: https://github.com/k2-fsa/sherpa-onnx

Original sherpa-onnx model repository:

csukuangfj/sherpa-onnx-whisper-tiny https://huggingface.co/csukuangfj/sherpa-onnx-whisper-tiny

License & Attribution

OpenAI Whisper is released under the MIT License.

This repository does not claim authorship of the original Whisper model architecture, weights, or sherpa-onnx conversion.

The original copyright and license terms of the upstream projects remain applicable.