Update README.md

README.md

@@ -20,7 +20,7 @@ https://github.com/akjava/Matcha-TTS-Japanese
 Matcha-TTS checkpoint - epoch seems big but train with only 290 audios
 ### ONNX
 
-onnx simplified 
+onnx simplified loading speed is now 1.5 times faster.
 ```
 from onnxsim import simplify
 import onnx
@@ -30,9 +30,25 @@ model_simp, check = simplify(model)
 
 onnx.save(model_simp, "en001_6399_T2_simplify.onnx")
 ```
+
+timesteps is default(5) ,small time steps ;The infer speed is somewhat faster, but the quality is lower.
+
+If you need original onnx do like official way
+```
+python -m matcha.onnx.export checkpoint_epoch=5699.ckpt en001_5699t2.onnx  --vocoder-name hifigan_T2_v1 --n-timesteps 5 --vocoder-checkpoint generator_v1
+python -m matcha.onnx.export checkpoint_epoch=5699.ckpt en001_5699.onnx  --vocoder-name hifigan_univ_v1 --n-timesteps 5 --vocoder-checkpoint g_02500000
+```
+
 - T2 means Vocoder is hifigan_T2_v1
 - Unif means Voder is hifigan_univ_v1
 
+you can quantize this onnx,but 3 times smaller, but 4-5 times slower,that why I did't include that.
+```
+from onnxruntime.quantization import quantize_dynamic, QuantType
+quantized_model = quantize_dynamic(src_model_path, dst_model_path, weight_type=QuantType.QUInt8)
+```
+
+
 To use onnx need something,I'll add sample code later
 ### Audio
 I cut with VAD tools and denoise with resemble-enhance