Document source revision and quantization/export provenance for Silero-VAD-INT8-ONNX

README.md

@@ -31,8 +31,8 @@ newly trained model.
 |---|---|
 | Upstream ONNX repo | [onnx-community/silero-vad](https://huggingface.co/onnx-community/silero-vad) |
 | Original model | [snakers4/silero-vad](https://github.com/snakers4/silero-vad) (GitHub) |
-| Source revision/hash | Not documented in this repo |
-| Quantization recipe | Not documented in this repo |
+| Source revision/hash | `e71cae966052b992a7eca6b17738916ce0eca4ec` ([onnx-community/silero-vad](https://huggingface.co/onnx-community/silero-vad)) |
+| Quantization recipe | Pre-quantized INT8 ONNX mirrored from `onnx-community/silero-vad` (`onnx/model_int8.onnx`) |
 
 ## Files
 
@@ -56,8 +56,7 @@ Silero VAD is language-agnostic and works across a wide range of languages.
 
 ## Precision and Packaging
 
-INT8 quantization. The exact method (static vs. dynamic, per-tensor vs. per-channel) is
-not documented in this repository. Some accuracy loss relative to FP16/FP32 is expected,
+INT8 quantization from the ONNX Community export (mirrored unchanged). Some accuracy loss relative to FP16/FP32 is expected,
 but Silero VAD is small and INT8 quality is generally acceptable for VAD. No parity
 benchmark is documented here.
 
@@ -65,7 +64,6 @@ benchmark is documented here.
 
 - Requires 16 kHz mono audio input; other sample rates must be resampled.
 - INT8 may reduce detection sensitivity at the margins (very quiet or very short speech).
-- Quantization recipe not documented; reproducibility is not guaranteed.
 - No repository-specific VAD quality benchmark is documented here.
 
 ## License