Spaces:
Running on Zero
Running on Zero
File size: 10,249 Bytes
03022ee | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 | import os
import torch
import numpy as np
import kaldiio
import librosa
import torchaudio
import torchaudio.compliance.kaldi as Kaldi
from torch.nn.utils.rnn import pad_sequence
import onnxruntime as ort
try:
from funcineforge.download.file import download_from_url
except:
print("urllib is not installed, if you infer from url, please install it first.")
import subprocess
from subprocess import CalledProcessError, run
def is_ffmpeg_installed():
try:
output = subprocess.check_output(["ffmpeg", "-version"], stderr=subprocess.STDOUT)
return "ffmpeg version" in output.decode("utf-8")
except (subprocess.CalledProcessError, FileNotFoundError):
return False
use_ffmpeg = False
if is_ffmpeg_installed():
use_ffmpeg = True
else:
print(
"Notice: ffmpeg is not installed. torchaudio is used to load audio\n"
"If you want to use ffmpeg backend to load audio, please install it by:"
"\n\tsudo apt install ffmpeg # ubuntu"
"\n\t# brew install ffmpeg # mac"
)
def load_audio_text_image_video(
data_or_path_or_list,
fs: int = 16000,
audio_fs: int = 16000,
data_type="sound",
tokenizer=None,
**kwargs,
):
if isinstance(data_or_path_or_list, (list, tuple)):
if data_type is not None and isinstance(data_type, (list, tuple)):
data_types = [data_type] * len(data_or_path_or_list)
data_or_path_or_list_ret = [[] for d in data_type]
for i, (data_type_i, data_or_path_or_list_i) in enumerate(
zip(data_types, data_or_path_or_list)
):
for j, (data_type_j, data_or_path_or_list_j) in enumerate(
zip(data_type_i, data_or_path_or_list_i)
):
data_or_path_or_list_j = load_audio_text_image_video(
data_or_path_or_list_j,
fs=fs,
audio_fs=audio_fs,
data_type=data_type_j,
tokenizer=tokenizer,
**kwargs,
)
data_or_path_or_list_ret[j].append(data_or_path_or_list_j)
return data_or_path_or_list_ret
else:
return [
load_audio_text_image_video(
audio, fs=fs, audio_fs=audio_fs, data_type=data_type, **kwargs
)
for audio in data_or_path_or_list
]
if isinstance(data_or_path_or_list, str) and data_or_path_or_list.startswith(
"http"
): # download url to local file
data_or_path_or_list = download_from_url(data_or_path_or_list)
if isinstance(data_or_path_or_list, str) and os.path.exists(data_or_path_or_list): # local file
if data_type is None or data_type in ["sound", "kaldi_ark_or_sound"]:
if kwargs.get("use_ffmpeg", False):
data_or_path_or_list = _load_audio_ffmpeg(data_or_path_or_list, sr=fs)
data_or_path_or_list = torch.from_numpy(
data_or_path_or_list
).squeeze() # [n_samples,]
else:
try:
data_or_path_or_list, audio_fs = torchaudio.load(data_or_path_or_list)
if kwargs.get("reduce_channels", True):
data_or_path_or_list = data_or_path_or_list.mean(0)
except:
data_or_path_or_list = _load_audio_ffmpeg(data_or_path_or_list, sr=fs)
data_or_path_or_list = torch.from_numpy(
data_or_path_or_list
).squeeze() # [n_samples,]
elif data_type == "text" and tokenizer is not None:
data_or_path_or_list = tokenizer.encode(data_or_path_or_list)
elif data_type == "image": # undo
pass
elif data_type == "video": # undo
pass
# if data_in is a file or url, set is_final=True
if "cache" in kwargs:
kwargs["cache"]["is_final"] = True
kwargs["cache"]["is_streaming_input"] = False
elif isinstance(data_or_path_or_list, str) and data_type == "text" and tokenizer is not None:
data_or_path_or_list = tokenizer.encode(data_or_path_or_list)
elif isinstance(data_or_path_or_list, np.ndarray): # audio sample point
data_or_path_or_list = torch.from_numpy(data_or_path_or_list).squeeze() # [n_samples,]
elif isinstance(data_or_path_or_list, str) and data_type in ["kaldi_ark", "kaldi_ark_or_sound", "sound"]:
if len(data_or_path_or_list.split()) == 2:
data_or_path_or_list, audio_fs = data_or_path_or_list.split()
audio_fs = int(audio_fs)
data_mat = kaldiio.load_mat(data_or_path_or_list)
if isinstance(data_mat, tuple):
audio_fs, mat = data_mat
else:
mat = data_mat
if mat.dtype == "int16":
mat = mat.astype(np.float32)
mat = mat / (2 ** 16)
elif mat.dtype == "int32":
mat = mat.astype(np.float32)
mat = mat / (2 ** 32)
if mat.ndim == 2:
mat = mat[:, 0]
data_or_path_or_list = torch.from_numpy(mat)
elif isinstance(data_or_path_or_list, bytes): # audio bytes
data_or_path_or_list = load_bytes(data_or_path_or_list)
else:
pass
print(f"unsupport data type: {data_or_path_or_list}, return raw data")
if audio_fs != fs and data_type != "text":
resampler = torchaudio.transforms.Resample(audio_fs, fs, dtype=data_or_path_or_list.dtype)
data_or_path_or_list = resampler(data_or_path_or_list[None, :])[0, :]
return data_or_path_or_list
class FBank(object):
def __init__(self,
n_mels,
sample_rate,
mean_nor: bool = False,
):
self.n_mels = n_mels
self.sample_rate = sample_rate
self.mean_nor = mean_nor
def __call__(self, wav, dither=0):
sr = 16000
assert sr == self.sample_rate
if len(wav.shape) == 1:
wav = wav.unsqueeze(0)
if wav.shape[0] > 1:
wav = torch.mean(wav, dim=0, keepdim=True)
assert len(wav.shape) == 2 and wav.shape[0] == 1, wav.shape
feat = Kaldi.fbank(wav, num_mel_bins=self.n_mels,
sample_frequency=sr, dither=dither)
# feat: [T, N]
if self.mean_nor:
feat = feat - feat.mean(0, keepdim=True)
return feat
class OnnxModel(object):
def __init__(self, pretrained_model):
session_options = ort.SessionOptions()
self.model = ort.InferenceSession(pretrained_model, session_options)
self.input_name = self.model.get_inputs()[0].name
self.output_name = self.model.get_outputs()[0].name
self.feature_extractor = FBank(n_mels=80, sample_rate=16000, mean_nor=True)
def __call__(self, wav):
feat = self.feature_extractor(torch.as_tensor(wav))
feat = feat.float().unsqueeze(0).numpy()
emb = self.model.run([self.output_name], {self.input_name: feat})[0]
return emb
def extract_campp_xvec(
wav_path: str = "",
target_sr: int = 16000,
**kwargs,
):
wav, sr = librosa.load(wav_path, dtype=np.float32, sr=target_sr, mono=True)
if sr != target_sr:
wav = librosa.resample(wav, orig_sr=sr, target_sr=target_sr)
onnx_path = kwargs.get("xvec_model", None)
model = OnnxModel(onnx_path)
xvec = model(wav)
return xvec
def load_bytes(input):
middle_data = np.frombuffer(input, dtype=np.int16)
middle_data = np.asarray(middle_data)
if middle_data.dtype.kind not in "iu":
raise TypeError("'middle_data' must be an array of integers")
dtype = np.dtype("float32")
if dtype.kind != "f":
raise TypeError("'dtype' must be a floating point type")
i = np.iinfo(middle_data.dtype)
abs_max = 2 ** (i.bits - 1)
offset = i.min + abs_max
array = np.frombuffer((middle_data.astype(dtype) - offset) / abs_max, dtype=np.float32)
return array
def extract_fbank(data, data_len=None, data_type: str = "sound", frontend=None, **kwargs):
if isinstance(data, np.ndarray):
data = torch.from_numpy(data)
if len(data.shape) < 2:
data = data[None, :] # data: [batch, N]
data_len = [data.shape[1]] if data_len is None else data_len
elif isinstance(data, torch.Tensor):
if len(data.shape) < 2:
data = data[None, :] # data: [batch, N]
data_len = [data.shape[1]] if data_len is None else data_len
elif isinstance(data, (list, tuple)):
data_list, data_len = [], []
for data_i in data:
if isinstance(data_i, np.ndarray):
data_i = torch.from_numpy(data_i)
data_list.append(data_i)
data_len.append(data_i.shape[0])
data = pad_sequence(data_list, batch_first=True) # data: [batch, N]
data, data_len = frontend(data, data_len, **kwargs)
if isinstance(data_len, (list, tuple)):
data_len = torch.tensor([data_len])
return data.to(torch.float32), data_len.to(torch.int32)
def _load_audio_ffmpeg(file: str, sr: int = 16000):
"""
Open an audio file and read as mono waveform, resampling as necessary
Parameters
----------
file: str
The audio file to open
sr: int
The sample rate to resample the audio if necessary
Returns
-------
A NumPy array containing the audio waveform, in float32 dtype.
"""
# This launches a subprocess to decode audio while down-mixing
# and resampling as necessary. Requires the ffmpeg CLI in PATH.
# fmt: off
cmd = [
"ffmpeg",
"-nostdin",
"-threads", "0",
"-i", file,
"-f", "s16le",
"-ac", "1",
"-acodec", "pcm_s16le",
"-ar", str(sr),
"-"
]
# fmt: on
try:
out = run(cmd, capture_output=True, check=True).stdout
except CalledProcessError as e:
raise RuntimeError(f"Failed to load audio: {e.stderr.decode()}") from e
return np.frombuffer(out, np.int16).flatten().astype(np.float32) / 32768.0
|