latentdivergence/whisper-tiny.en-onnx

whisper-tiny.en (ONNX, quantized) — FixFlix build

A pinned copy of the quantized ONNX export of openai/whisper-tiny.en, used by the FixFlix Android app for on-device caption generation.

This mirror exists so the app does not depend on a third-party repository staying available or unchanged. The weights are unmodified.

Contents

File	Size	Purpose
onnx/encoder_model_quantized.onnx	9.7 MB	Audio encoder
onnx/decoder_model_merged_quantized.onnx	29.3 MB	Text decoder with merged KV cache
tokenizer.json	2.1 MB	BPE tokenizer
config.json, generation_config.json, preprocessor_config.json	small	Model and decoding parameters

Total ≈ 41 MB. The app downloads these on first caption run rather than bundling them, and caches them in app storage.

Expected input

16 kHz mono audio, converted to an 80-bin log-mel spectrogram (25 ms window, 10 ms hop) — Whisper's standard preprocessing. Feeding a different sample rate produces fluent but incorrect transcription rather than an error.

Attribution and licence

• Base model: openai/whisper-tiny.en (MIT)
• ONNX export quantized by Xenova/whisper-tiny.en (Apache-2.0)

Redistributed under Apache-2.0. No changes were made to the model weights.