VCoklat/edgespeech-whisper-tiny-int8

Whisper-tiny INT8 ONNX

Dynamic INT8 quantized weights for openai/whisper-tiny, optimized for low-latency, edge-native speech recognition directly in the browser via ONNX Runtime WebAssembly.

Model Summary

• Base Model: openai/whisper-tiny
• Quantization Method: Dynamic INT8 (QuantType.QUInt8)
• Total Repository Size: ~103.4 MB
• Active Runtime Footprint: ~56 MB (Only loads Encoder + KV-cache Decoder during streaming inference)
• Intended Use: Client-side ASR execution for edgespeech-wasm

Files & Sizes

• encoder_model_int8.onnx (~9.6 MB): Audio feature extractor & encoder
• decoder_model_int8.onnx (~47.5 MB): Standard autoregressive text decoder
• decoder_with_past_model_int8.onnx (~46.3 MB): Decoder optimized with KV-cache for low-latency streaming