Download scripts/score_dir.py from melser/Reverb-TTS-onnx: direct link, hf CLI and curl.
- Browser
- Download file 1.66 kB
-
https://huggingface.co/melser/Reverb-TTS-onnx/resolve/main/scripts/score_dir.py
- Command line
-
hf download hf://melser/Reverb-TTS-onnx/scripts/score_dir.py
-
curl -L -o score_dir.py https://huggingface.co/melser/Reverb-TTS-onnx/resolve/main/scripts/score_dir.py
1.66 kB
| import sys, os, re, json, unicodedata, glob | |
| import soundfile as sf | |
| from faster_whisper import WhisperModel | |
| texts=["Buenos días, ¿cómo estás hoy? Me llamo Ana y voy a contarte una historia sobre el mercado.", | |
| "¡Qué sorpresa verte aquí! No esperaba encontrarte en la biblioteca esta tarde.", | |
| "El tren sale a las ocho de la mañana, así que debemos levantarnos temprano.", | |
| "Uno, dos, tres, cuatro, cinco, seis, siete, ocho, nueve, diez."] | |
| def norm(s): | |
| s=unicodedata.normalize('NFKD',s.lower()); s=''.join(c for c in s if not unicodedata.combining(c)) | |
| for a,b in [('10','diez'),('1','uno'),('2','dos'),('3','tres'),('4','cuatro'),('5','cinco'),('6','seis'),('7','siete'),('8','ocho'),('9','nueve')]: s=re.sub(rf'\b{a}\b',b,s) | |
| return re.findall(r"[a-zñ]+",s) | |
| def ed(a,b): | |
| d=list(range(len(b)+1)) | |
| for i,x in enumerate(a,1): | |
| p,d[0]=d[0],i | |
| for j,y in enumerate(b,1): p,d[j]=d[j],min(d[j]+1,d[j-1]+1,p+(x!=y)) | |
| return d[-1] | |
| m=WhisperModel('small',device='cpu',compute_type='int8',cpu_threads=6) | |
| for d in sys.argv[1:]: | |
| errs=words=bad=n=0; dur=0 | |
| for ti,t in enumerate(texts): | |
| for f in sorted(glob.glob(f"{d}/t{ti}_s*.wav")): | |
| a,sr=sf.read(f,dtype='float32'); dur+=len(a)/sr | |
| segs,_=m.transcribe(a if sr==16000 else __import__('scipy.signal',fromlist=['x']).resample_poly(a,16000,sr).astype('float32'),language='es',beam_size=5) | |
| hyp=' '.join(x.text for x in segs); ref=norm(t); e=ed(ref,norm(hyp)); errs+=e; words+=len(ref); n+=1; bad+= e/len(ref)>0.5 | |
| print(json.dumps({"set":d,"WER":round(errs/words,3),"bad(>50%)":f"{bad}/{n}","avg_audio_s":round(dur/n,2)}),flush=True) | |