From eb1574b813e230b156fc09eaaf03227b1b0b4134 Mon Sep 17 00:00:00 2001
From: weilikai <jasper@talkus.fun>
Date: 星期六, 20 九月 2025 22:41:05 +0800
Subject: [PATCH] fix: support loading .pcm (16k 1c 16bit) audio files in load_utils.py (#2667) (#2668)
---
funasr/utils/load_utils.py | 229 ++++++++++++++++++++++++++++++++++++++++++++++++--------
1 files changed, 194 insertions(+), 35 deletions(-)
diff --git a/funasr/utils/load_utils.py b/funasr/utils/load_utils.py
index 7748172..d208f7d 100644
--- a/funasr/utils/load_utils.py
+++ b/funasr/utils/load_utils.py
@@ -1,6 +1,7 @@
import os
import torch
import json
+from io import BytesIO
import torch.distributed as dist
import numpy as np
import kaldiio
@@ -9,45 +10,110 @@
import time
import logging
from torch.nn.utils.rnn import pad_sequence
+
try:
from funasr.download.file import download_from_url
except:
print("urllib is not installed, if you infer from url, please install it first.")
+import pdb
+import subprocess
+from subprocess import CalledProcessError, run
+
+try:
+ from pydub import AudioSegment
+except:
+ pass
+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
-def load_audio_text_image_video(data_or_path_or_list, fs: int = 16000, audio_fs: int = 16000, data_type="sound", tokenizer=None, **kwargs):
+
+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)
+ 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
+ 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://", "https://")
+ ): # 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 (isinstance(data_or_path_or_list, str) and os.path.exists(data_or_path_or_list)) or hasattr(data_or_path_or_list, 'read'): # local file or bytes io
if data_type is None or data_type == "sound":
- 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)
+ if hasattr(data_or_path_or_list, "read") and hasattr(data_or_path_or_list, "seek"):
+ data_or_path_or_list.seek(0)
+ # if use_ffmpeg:
+ # 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:
+ # 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)
+ 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
+ with open(data_or_path_or_list, "r") as f:
+ data_or_path_or_list = tokenizer.encode(f.read().strip())
+ elif data_type == "image": # undo
pass
- elif data_type == "video": # undo
+ 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
@@ -55,42 +121,89 @@
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,]
+ 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 == "kaldi_ark":
+ 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" or mat.dtype == "int32":
+ mat = mat.astype(np.float64)
+ mat = mat / 32768
+ if mat.ndim == 2:
+ mat = mat[:, 0]
+ data_or_path_or_list = mat
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)
data_or_path_or_list = resampler(data_or_path_or_list[None, :])[0, :]
return data_or_path_or_list
+
def load_bytes(input):
+ try:
+ input = validate_frame_rate(input)
+ except:
+ pass
middle_data = np.frombuffer(input, dtype=np.int16)
middle_data = np.asarray(middle_data)
- if middle_data.dtype.kind not in 'iu':
+ 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':
+ 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):
- # import pdb;
- # pdb.set_trace()
+
+def validate_frame_rate(
+ input,
+ fs: int = 16000,
+):
+
+ # 灏嗘枃浠惰鍙栦负瀛楄妭娴�
+ byte_data = BytesIO(input)
+
+ # 浣跨敤 pydub 鍔犺浇闊抽
+ try:
+ audio = AudioSegment.from_file(byte_data)
+ except:
+ raise RuntimeError(
+ "You are decoding the pcm data, please install pydub first. via `pip install pydub`."
+ )
+
+ # 纭繚閲囨牱鐜囦负 16000 Hz
+ if audio.frame_rate != fs:
+ audio = audio.set_frame_rate(fs)
+
+ # 灏嗛噸鏂伴噰鏍峰悗鐨勯煶棰戝鍑轰负瀛楄妭娴�
+ output = BytesIO()
+ audio.export(output, format="wav")
+ output.seek(0)
+
+ # 鑾峰彇閲嶆柊閲囨牱鍚庣殑瀛楄妭娴佹暟鎹�
+ input = output.read()
+
+ return input
+
+
+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 = 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 = 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 = [], []
@@ -99,13 +212,59 @@
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]
- # import pdb;
- # pdb.set_trace()
- # if data_type == "sound":
+ 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
+ pcm_params = []
+ if file.lower().endswith('.pcm'):
+ pcm_params = [
+ "-f", "s16le",
+ "-ar", str(sr),
+ "-ac", "1"
+ ]
+
+ cmd = [
+ "ffmpeg",
+ "-nostdin",
+ "-threads", "0",
+ *pcm_params, # PCM files need input format specified before -i since PCM is raw data without headers
+ "-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
--
Gitblit v1.9.1