mrq/vall-e
5
9
Fork 0
Code Issues 8 Pull Requests Packages Projects Releases Wiki Activity

ugh

Browse Source
This commit is contained in:
mrq 2025-02-09 12:30:55 -06:00
parent 953015748f
commit 8e7d19105c
3 changed files with 51 additions and 40 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

73
vall_e/emb/similar.py
View File

@ -111,9 +111,7 @@ def speaker_similarity_embedding(
def batch_similar_utterances(
speaker_path,
yaml,
text=False,
audio_backend="encodec",
device="cuda",
dtype="float16",
amp=False,
@ -127,7 +125,9 @@ def batch_similar_utterances(
min_duration=0,
max_duration=0,
storage_backend="slop"
audio_backend="encodec",
storage_backend="slop",
similarity_backend="resp",
):
global tts
@ -156,7 +156,7 @@ def batch_similar_utterances(
extension = filename.split(".")[-1]
filename = filename.replace(f".{extension}", "")
if text:
if similarity_backend == "text":
if extension not in artifact_extension:
raise Exception("!")
@ -183,34 +183,37 @@ def batch_similar_utterances(
phn = phn.replace(f"({metadata['language']})", "")
embedding = tts.text_embedding( phn )
else:
elif similarity_backend == "resp":
# treat embeddings as features, if provided quantized audio
if extension in artifact_extension:
artifact = np.load(f'./{speaker_path}/{filename}.{extension}', allow_pickle=True)[()]
duration = artifact["metadata"]["original_length"] / artifact["metadata"]["sample_rate"]
if extension not in artifact_extension:
continue
artifact = np.load(f'./{speaker_path}/{filename}.{extension}', allow_pickle=True)[()]
duration = artifact["metadata"]["original_length"] / artifact["metadata"]["sample_rate"]
"""
if 0 < min_duration and duration < min_duration:
continue
if 0 < max_duration and max_duration < duration:
continue
"""
"""
if 0 < min_duration and duration < min_duration:
continue
if 0 < max_duration and max_duration < duration:
continue
"""
qnt = torch.from_numpy(artifact["codes"].astype(int))[0].t().to(dtype=torch.int16, device=device)
qnt = torch.from_numpy(artifact["codes"].astype(int))[0].t().to(dtype=torch.int16, device=device)
if trim_duration > 0:
qnt = trim( qnt, int( cfg.dataset.frames_per_second * trim_duration ) )
embedding = tts.audio_embedding( qnt )
# try and extract features from the raw audio itself
if trim_duration > 0:
qnt = trim( qnt, int( cfg.dataset.frames_per_second * trim_duration ) )
embedding = tts.audio_embedding( qnt )
# try and extract features from the raw audio itself
else:
# qnt = tts.encode_audio(f'./{speaker_path}/{filename}', trim_length=3.0).to(device)
if similarity_backend == "wavlm":
embedding = speaker_similarity_embedding( f'./{speaker_path}/{filename}.{extension}' )
else:
# qnt = tts.encode_audio(f'./{speaker_path}/{filename}', trim_length=3.0).to(device)
wav, sr = load_audio( f'./{speaker_path}/{filename}.{extension}' )
duration = wav.shape[-1] / sr
"""
duration = wav.shape[-1] / sr
if 0 < min_duration and duration < min_duration:
continue
@ -223,6 +226,7 @@ def batch_similar_utterances(
embedding = mfcc(wav.to(device))[0].t()
dim_shape = embedding.shape[-1]
if slop:
embedding = embedding.sum(dim=0)
@ -254,10 +258,10 @@ def batch_similar_utterances(
top_k = min( top_k, len(keys) )
if top_k == 0:
return
top_k = len(keys)
# fill any missing keys with a null embedding to keep the order the same
null_embedding = torch.zeros( (1024,), device=tts.device, dtype=tts.dtype )
null_embedding = torch.zeros( (dim_shape,), device=tts.device, dtype=tts.dtype )
embeddings = torch.stack( [ feature if feature is not None else null_embedding for feature in features.values() ] )
sorted_similarities = {}
@ -292,17 +296,18 @@ def main():
parser.add_argument("--use-dataset", action="store_true")
parser.add_argument("--yaml", type=Path)
parser.add_argument("--text", action="store_true")
# dropped, because this might mess with the indices to map to
"""
parser.add_argument("--trim-duration", type=float, default=3.0)
parser.add_argument("--min-duration", type=float, default=0)
parser.add_argument("--max-duration", type=float, default=0)
"""
parser.add_argument("--storage-backend", type=str, default="slop")
parser.add_argument("--top-k", type=int, default=8)
parser.add_argument("--storage-backend", type=str, default="slop")
parser.add_argument("--similarity-backend", type=str, default="resp")
parser.add_argument("--audio-backend", type=str, default="encodec")
parser.add_argument("--dtype", type=str, default="float16")
parser.add_argument("--amp", action="store_true")
parser.add_argument("--device", type=str, default="cuda")
@ -336,15 +341,16 @@ def main():
similarities = batch_similar_utterances(
speaker_path=cfg.data_dir / speaker_name,
yaml=args.yaml,
text=args.text,
top_k=args.top_k,
#trim_duration=args.trim_duration,
#min_duration=args.min_duration,
#max_duration=args.max_duration,
audio_backend=args.audio_backend,
storage_backend=args.storage_backend,
similarity_backend=args.similarity_backend,
metadata_keys=metadata_keys,
audio_backend=args.audio_backend,
device=args.device,
dtype=args.dtype,
amp=args.amp,
@ -384,22 +390,25 @@ def main():
similarities = batch_similar_utterances(
speaker_path=args.input_speaker,
yaml=args.yaml,
text=args.text,
top_k=args.top_k,
#trim_duration=args.trim_duration,
#min_duration=args.min_duration,
#max_duration=args.max_duration,
audio_backend=args.audio_backend,
device=args.device,
dtype=args.dtype,
amp=args.amp,
audio_backend=args.audio_backend,
storage_backend=args.storage_backend,
similarity_backend=args.similarity_backend,
verbose=True,
)
json_write( similarities, "similarities.json" )
# and print
for filename, sim in similarities.items():
print(f'{filename}: {sim}')

14
vall_e/emb/transcribe.py
View File

@ -165,11 +165,17 @@ def transcribe(
start = 0
end = 0
segments = []
for segment in result["chunks"]:
info = torchaudio.info(audio)
duration = info.num_frames / info.sample_rate
for segment in result["chunks"]:
text = segment["text"]
if "timestamp" in segment:
s, e = segment["timestamp"]
if not e:
e = duration
start = min( start, s )
end = max( end, e )
else:
@ -285,12 +291,12 @@ def transcribe_batch(
if not os.path.isdir(f'./{input_audio}/{dataset_name}/'):
continue
if group_name in ignore_groups:
if dataset_name in ignore_groups:
continue
if only_groups and group_name not in only_groups:
if only_groups and dataset_name not in only_groups:
continue
for speaker_id in tqdm(process_items(os.listdir(f'./{input_audio}/{dataset_name}/')), desc="Processing speaker"):
for speaker_id in tqdm(process_items(os.listdir(f'./{input_audio}/{dataset_name}/'), stride=stride, stride_offset=stride_offset), desc="Processing speaker"):
if not os.path.isdir(f'./{input_audio}/{dataset_name}/{speaker_id}'):
continue

4
vall_e/models/arch/__init__.py
View File

@ -46,15 +46,11 @@ except Exception as e:
ERROR_ARCHES["bitnet"] = e
pass
from .mixtral import MixtralModel, MixtralConfig, MixtralAttention, MixtralAttention_Adapted, MixtralModel_Adapted, load_balancing_loss_func
AVAILABLE_ARCHES.append("mixtral")
"""
try:
from .mixtral import MixtralModel, MixtralConfig, MixtralAttention, MixtralAttention_Adapted, MixtralModel_Adapted, load_balancing_loss_func
AVAILABLE_ARCHES.append("mixtral")
except Exception as e:
ERROR_ARCHES["mixtral"] = e
"""
try:
from .mamba import MambaModel, Mamba2Model, MambaConfig, Mamba2Config