added dataset transcription helper script (now I don't ever have to touch ai-voice-cloning) (to-do: unify scripts into the module)

This commit is contained in:
mrq 2024-04-21 17:43:20 -05:00
parent b251669536
commit ffc334cf58
3 changed files with 82 additions and 6 deletions

View File

@ -0,0 +1,79 @@
import os
import json
import torch
import torchaudio
import whisperx
from tqdm.auto import tqdm
from pathlib import Path
device = "cuda"
batch_size = 16
dtype = "float16"
model_size = "large-v2"
input_audio = "voice"
output_dataset = "metadata"
skip_existing = True
model = whisperx.load_model(model_size, device, compute_type=dtype)
align_model, align_model_metadata, align_model_language = (None, None, None)
def pad(num, zeroes):
return str(num).zfill(zeroes+1)
for dataset_name in os.listdir(f'./{input_audio}/'):
if not os.path.isdir(f'./{input_audio}/{dataset_name}/'):
print("Is not dir:", f'./{input_audio}/{dataset_name}/')
continue
for speaker_id in tqdm(os.listdir(f'./{input_audio}/{dataset_name}/'), desc="Processing speaker"):
if not os.path.isdir(f'./{input_audio}/{dataset_name}/{speaker_id}'):
print("Is not dir:", f'./{input_audio}/{dataset_name}/{speaker_id}')
continue
outpath = Path(f'./{output_dataset}/{dataset_name}/{speaker_id}/whisper.json')
if outpath.exists():
metadata = json.loads(open(outpath, 'r', encoding='utf-8').read())
else:
os.makedirs(f'./{output_dataset}/{dataset_name}/{speaker_id}/', exist_ok=True)
metadata = {}
for filename in tqdm(os.listdir(f'./{input_audio}/{dataset_name}/{speaker_id}/'), desc=f"Processing speaker: {speaker_id}"):
if skip_existing and filename in metadata:
continue
inpath = f'./{input_audio}/{dataset_name}/{speaker_id}/{filename}'
metadata[filename] = {
"segments": [],
"language": "",
"text": [],
}
audio = whisperx.load_audio(inpath)
result = model.transcribe(audio, batch_size=batch_size)
language = result["language"]
if align_model_language != language:
tqdm.write(f'Loading language: {language}')
align_model, align_model_metadata = whisperx.load_align_model(language_code=language, device=device)
align_model_language = language
result = whisperx.align(result["segments"], align_model, align_model_metadata, audio, device, return_char_alignments=False)
metadata[filename]["segments"] = result["segments"]
metadata[filename]["language"] = language
text = []
for segment in result["segments"]:
id = len(text)
text.append( segment["text"] )
metadata[filename]["segments"][id]["id"] = id
metadata[filename]["text"] = " ".join(text).strip()
open(outpath, 'w', encoding='utf-8').write(json.dumps(metadata))

View File

@ -177,10 +177,8 @@ def _get_phones(path, language="en"):
content = metadata["phonemes"]
else:
content = open(_get_phone_path(path), "r", encoding="utf-8").read().split(" ")
content = _cleanup_phones( content )
return "".join(content)
#return ["<s>"] + [ " " if not p else p for p in content ] + ["</s>"]
def _interleaved_reorder(l, fn):
groups = defaultdict(list)
@ -431,8 +429,6 @@ class Dataset(_Dataset):
text = cfg.hdf5[key]["text"][:]
resps = cfg.hdf5[key]["audio"][:, :]
text = np.array( _cleanup_phones( text, targets=[ self.phone_symmap[" "] ] ) )
text = torch.from_numpy(text).to(self.text_dtype)
resps = torch.from_numpy(resps).to(torch.int16)
@ -455,12 +451,13 @@ class Dataset(_Dataset):
txt = cfg.hdf5[key]["text"][:]
qnt = cfg.hdf5[key]["audio"][:, :]
txt = np.array( _cleanup_phones( txt, targets=[ self.phone_symmap[" "] ] ) )
txt = np.array( txt )
txt = torch.from_numpy(txt).to(self.text_dtype)
qnt = torch.from_numpy(qnt).to(torch.int16)
else:
txt = torch.tensor([*map(self.phone_symmap.get, _get_phones(sampled_path))]).to(self.text_dtype)
#txt = torch.tensor([*map(self.phone_symmap.get, _get_phones(sampled_path))]).to(self.text_dtype)
txt = torch.tensor(tokenize(_get_phones(sampled_path))).to(self.text_dtype)
qnt = _load_quants(sampled_path)
# <s>[original text] [new text]</s>