From 2cbebbb33435dc490b6a092f2e5903c7f0e3e33c Mon Sep 17 00:00:00 2001
From: zhifu gao <zhifu.gzf@alibaba-inc.com>
Date: 星期二, 31 一月 2023 17:54:02 +0800
Subject: [PATCH] Merge pull request #52 from alibaba-damo-academy/dev

---
 funasr/datasets/iterable_dataset.py |   16 +++++++++++++++-
 funasr/utils/asr_utils.py           |   29 ++++++++++++++++++++++-------
 2 files changed, 37 insertions(+), 8 deletions(-)

diff --git a/funasr/datasets/iterable_dataset.py b/funasr/datasets/iterable_dataset.py
index bed295b..1fc9270 100644
--- a/funasr/datasets/iterable_dataset.py
+++ b/funasr/datasets/iterable_dataset.py
@@ -13,12 +13,15 @@
 import numpy as np
 import soundfile
 import torch
+import torchaudio
 from torch.utils.data.dataset import IterableDataset
 from typeguard import check_argument_types
 import os.path
 
 from funasr.datasets.dataset import ESPnetDataset
 
+
+SUPPORT_AUDIO_TYPE_SETS = ['flac', 'mp3', 'm4a', 'ogg', 'opus', 'wav', 'wma']
 
 def load_kaldi(input):
     retval = kaldiio.load_mat(input)
@@ -60,7 +63,7 @@
 
 
 DATA_TYPES = {
-    "sound": lambda x: soundfile.read(x)[0],
+    "sound": lambda x: torchaudio.load(x)[0][0].numpy(),
     "kaldi_ark": load_kaldi,
     "bytes": load_bytes,
     "waveform": lambda x: x,
@@ -201,6 +204,11 @@
             uid = os.path.basename(self.path_name_type_list[0][0]).split(".")[0]
             name = self.path_name_type_list[0][1]
             _type = self.path_name_type_list[0][2]
+            if _type == "sound":
+                audio_type = os.path.basename(value).split(".")[1].lower()
+                if audio_type not in SUPPORT_AUDIO_TYPE_SETS:
+                    raise NotImplementedError(
+                        f'Not supported audio type: {audio_type}')
             func = DATA_TYPES[_type]
             array = func(value)
             data[name] = array
@@ -286,6 +294,11 @@
                 data = {}
                 # 2.a. Load data streamingly
                 for value, (path, name, _type) in zip(values, self.path_name_type_list):
+                    if _type == "sound":
+                        audio_type = os.path.basename(value).split(".")[1].lower()
+                        if audio_type not in SUPPORT_AUDIO_TYPE_SETS:
+                            raise NotImplementedError(
+                                f'Not supported audio type: {audio_type}')
                     func = DATA_TYPES[_type]
                     # Load entry
                     array = func(value)
@@ -322,3 +335,4 @@
 
         if count == 0:
             raise RuntimeError("No iteration")
+
diff --git a/funasr/utils/asr_utils.py b/funasr/utils/asr_utils.py
index 0bf903c..aa5c9db 100644
--- a/funasr/utils/asr_utils.py
+++ b/funasr/utils/asr_utils.py
@@ -4,7 +4,7 @@
 import struct
 from typing import Any, Dict, List, Union
 
-import librosa
+import torchaudio
 import numpy as np
 import pkg_resources
 from modelscope.utils.logger import get_logger
@@ -18,6 +18,7 @@
 
 global_asr_language = 'zh-cn'
 
+SUPPORT_AUDIO_TYPE_SETS = ['flac', 'mp3', 'm4a', 'ogg', 'opus', 'wav', 'wma']
 
 def get_version():
     return float(pkg_resources.get_distribution('easyasr').version)
@@ -57,12 +58,16 @@
     if r_recog_type is None and audio_in is not None:
         # audio_in is wav, recog_type is wav_file
         if os.path.isfile(audio_in):
-            if audio_in.endswith('.wav') or audio_in.endswith('.WAV'):
+            audio_type = os.path.basename(audio_in).split(".")[1].lower()
+            if audio_type in SUPPORT_AUDIO_TYPE_SETS:
                 r_recog_type = 'wav'
                 r_audio_format = 'wav'
-            elif audio_in.endswith('.scp') or audio_in.endswith('.SCP'):
+            elif audio_type == "scp":
                 r_recog_type = 'wav'
                 r_audio_format = 'scp'
+            else:
+                raise NotImplementedError(
+                    f'Not supported audio type: {audio_type}')
 
         # recog_type is datasets_file
         elif os.path.isdir(audio_in):
@@ -123,14 +128,15 @@
 def get_sr_from_wav(fname: str):
     fs = None
     if os.path.isfile(fname):
-        audio, fs = librosa.load(fname, sr=None)
+        audio, fs = torchaudio.load(fname)
         return fs
     elif os.path.isdir(fname):
         dir_files = os.listdir(fname)
         for file in dir_files:
             file_path = os.path.join(fname, file)
             if os.path.isfile(file_path):
-                if file_path.endswith('.wav') or file_path.endswith('.WAV'):
+                audio_type = os.path.basename(file_path).split(".")[1].lower()
+                if audio_type in SUPPORT_AUDIO_TYPE_SETS:
                     fs = get_sr_from_wav(file_path)
             elif os.path.isdir(file_path):
                 fs = get_sr_from_wav(file_path)
@@ -146,7 +152,14 @@
     for file in dir_files:
         file_path = os.path.join(dir_path, file)
         if os.path.isfile(file_path):
-            if file_path.endswith(ends):
+            if ends == ".wav" or ends == ".WAV":
+                audio_type = os.path.basename(file_path).split(".")[1].lower()
+                if audio_type in SUPPORT_AUDIO_TYPE_SETS:
+                    return True
+                else:
+                    raise NotImplementedError(
+                        f'Not supported audio type: {audio_type}')
+            elif file_path.endswith(ends):
                 return True
         elif os.path.isdir(file_path):
             if find_file_by_ends(file_path, ends):
@@ -160,7 +173,8 @@
     for file in dir_files:
         file_path = os.path.join(dir_path, file)
         if os.path.isfile(file_path):
-            if file_path.endswith('.wav') or file_path.endswith('.WAV'):
+            audio_type = os.path.basename(file_path).split(".")[1].lower()
+            if audio_type in SUPPORT_AUDIO_TYPE_SETS:
                 wav_list.append(file_path)
         elif os.path.isdir(file_path):
             recursion_dir_all_wav(wav_list, file_path)
@@ -333,3 +347,4 @@
         percent = 1
     res = int(50 * percent) * '#'
     print('\r[%-50s] %d%%' % (res, int(100 * percent)), end='')
+

--
Gitblit v1.9.1