Download scripts/parity_codec.py from melser/Reverb-TTS-onnx: direct link, hf CLI and curl.
- Browser
- Download file 1.68 kB
-
https://huggingface.co/melser/Reverb-TTS-onnx/resolve/main/scripts/parity_codec.py
- Command line
-
hf download hf://melser/Reverb-TTS-onnx/scripts/parity_codec.py
-
curl -L -o parity_codec.py https://huggingface.co/melser/Reverb-TTS-onnx/resolve/main/scripts/parity_codec.py
1.68 kB
| import os | |
| REF_WAV = os.environ.get("REF_WAV", "reference.wav") # clean 5-10 s speech, 24 kHz or any rate | |
| REF_TEXT = os.environ.get("REF_TEXT", "") # exact transcript of REF_WAV (gen_eval.py only) | |
| import torch, numpy as np, onnxruntime as ort, time | |
| import torch.nn.functional as F, torchaudio, scipy.signal as ss, soundfile as sf | |
| from transformers import HiggsAudioV2TokenizerModel | |
| m=HiggsAudioV2TokenizerModel.from_pretrained("eustlb/higgs-audio-v2-tokenizer").eval() | |
| so=ort.SessionOptions(); so.intra_op_num_threads=6 | |
| dec=ort.InferenceSession("onnx_fp16/codec_decoder.onnx",so,providers=["CPUExecutionProvider"]) | |
| enc=ort.InferenceSession("onnx_fp16/codec_encoder.onnx",so,providers=["CPUExecutionProvider"]) | |
| a,sr=sf.read(REF_WAV); a=torch.tensor(a,dtype=torch.float32) | |
| if sr!=24000: a=torchaudio.functional.resample(a,sr,24000) | |
| for secs in (7.0,3.0,1.4): | |
| x=a[:int(secs*24000)]; x=F.pad(x,(0,-x.numel()%960)); w24=x[None,None] | |
| with torch.inference_mode(): ref=m.encode(w24).audio_codes; refwav=m.decode(ref).audio_values.numpy() | |
| w16=torchaudio.functional.resample(w24,24000,16000)[:,0].numpy() | |
| t=time.time(); c1=enc.run(None,{"wav24":w24.numpy(),"wav16":w16})[0]; te=time.time()-t | |
| w16s=ss.resample_poly(w24[0,0].numpy().astype(np.float64),2,3).astype(np.float32)[None] | |
| c2=enc.run(None,{"wav24":w24.numpy(),"wav16":w16s})[0] | |
| t=time.time(); au=dec.run(None,{"codes":ref.numpy()})[0]; td=time.time()-t | |
| print(f"{secs}s: frames {ref.shape[2]} | encoder(torchaudio rs) agree {np.mean(c1==ref.numpy())*100:.2f}% ({te:.1f}s) | encoder(scipy rs) agree {np.mean(c2==ref.numpy())*100:.2f}% | decoder max diff {np.abs(au-refwav).max():.2e} ({td:.2f}s)") | |