File size: 4,231 Bytes
5fc76ef |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 |
import io
from hashlib import sha256
from pathlib import Path
from typing import Callable, Literal, Tuple
import torch
import torchaudio
from loguru import logger
from fish_speech.models.vqgan.modules.firefly import FireflyArchitecture
from fish_speech.utils.file import (
AUDIO_EXTENSIONS,
audio_to_bytes,
list_files,
read_ref_text,
)
from fish_speech.utils.schema import ServeReferenceAudio
class ReferenceLoader:
def __init__(self) -> None:
"""
Component of the TTSInferenceEngine class.
Loads and manages the cache for the reference audio and text.
"""
self.ref_by_id: dict = {}
self.ref_by_hash: dict = {}
# Make Pylance happy (attribut/method not defined...)
self.decoder_model: FireflyArchitecture
self.encode_reference: Callable
# Define the torchaudio backend
backends = torchaudio.list_audio_backends()
if "ffmpeg" in backends:
self.backend = "ffmpeg"
else:
self.backend = "soundfile"
def load_by_id(
self,
id: str,
use_cache: Literal["on", "off"],
) -> Tuple:
# Load the references audio and text by id
ref_folder = Path("references") / id
ref_folder.mkdir(parents=True, exist_ok=True)
ref_audios = list_files(
ref_folder, AUDIO_EXTENSIONS, recursive=True, sort=False
)
if use_cache == "off" or id not in self.ref_by_id:
# If the references are not already loaded, encode them
prompt_tokens = [
self.encode_reference(
# decoder_model=self.decoder_model,
reference_audio=audio_to_bytes(str(ref_audio)),
enable_reference_audio=True,
)
for ref_audio in ref_audios
]
prompt_texts = [
read_ref_text(str(ref_audio.with_suffix(".lab")))
for ref_audio in ref_audios
]
self.ref_by_id[id] = (prompt_tokens, prompt_texts)
else:
# Reuse already encoded references
logger.info("Use same references")
prompt_tokens, prompt_texts = self.ref_by_id[id]
return prompt_tokens, prompt_texts
def load_by_hash(
self,
references: list[ServeReferenceAudio],
use_cache: Literal["on", "off"],
) -> Tuple:
# Load the references audio and text by hash
audio_hashes = [sha256(ref.audio).hexdigest() for ref in references]
cache_used = False
prompt_tokens, prompt_texts = [], []
for i, ref in enumerate(references):
if use_cache == "off" or audio_hashes[i] not in self.ref_by_hash:
# If the references are not already loaded, encode them
prompt_tokens.append(
self.encode_reference(
reference_audio=ref.audio,
enable_reference_audio=True,
)
)
prompt_texts.append(ref.text)
self.ref_by_hash[audio_hashes[i]] = (prompt_tokens, prompt_texts)
else:
# Reuse already encoded references
prompt_tokens, prompt_texts = self.ref_by_hash[audio_hashes[i]]
cache_used = True
if cache_used:
logger.info("Use same references")
return prompt_tokens, prompt_texts
def load_audio(self, reference_audio, sr):
"""
Load the audio data from a file or bytes.
"""
if len(reference_audio) > 255 or not Path(reference_audio).exists():
audio_data = reference_audio
reference_audio = io.BytesIO(audio_data)
waveform, original_sr = torchaudio.load(reference_audio, backend=self.backend)
if waveform.shape[0] > 1:
waveform = torch.mean(waveform, dim=0, keepdim=True)
if original_sr != sr:
resampler = torchaudio.transforms.Resample(
orig_freq=original_sr, new_freq=sr
)
waveform = resampler(waveform)
audio = waveform.squeeze().numpy()
return audio
|