Spaces:
Running
Running
Upload 5 files
Browse files- Dockerfile +3 -7
- app.py +4 -77
- process.py +81 -0
Dockerfile
CHANGED
@@ -12,11 +12,7 @@ RUN apt-get update && \
|
|
12 |
RUN python3 -m pip install --upgrade pip
|
13 |
|
14 |
WORKDIR /app
|
15 |
-
COPY requirements.txt .
|
16 |
|
17 |
-
# requirements.txt
|
18 |
-
|
19 |
-
|
20 |
-
COPY . .
|
21 |
-
|
22 |
-
CMD ["python3", "app.py"]
|
|
|
12 |
RUN python3 -m pip install --upgrade pip
|
13 |
|
14 |
WORKDIR /app
|
|
|
15 |
|
16 |
+
# requirements.txt をコンテナ内にコピーして、必要なパッケージをインストール
|
17 |
+
COPY requirements.txt /app/
|
18 |
+
RUN pip install --no-cache-dir -r requirements.txt
|
|
|
|
|
|
app.py
CHANGED
@@ -1,84 +1,11 @@
|
|
1 |
from flask import Flask, request, jsonify, render_template, send_from_directory
|
2 |
import base64
|
|
|
3 |
import os
|
4 |
import shutil
|
5 |
-
|
6 |
-
import string
|
7 |
-
import random
|
8 |
-
from datetime import datetime
|
9 |
-
from pyannote.audio import Model, Inference
|
10 |
-
from pydub import AudioSegment
|
11 |
-
|
12 |
-
# Hugging Face のトークン取得(環境変数 HF に設定)
|
13 |
-
hf_token = os.environ.get("HF")
|
14 |
-
if hf_token is None:
|
15 |
-
raise ValueError("HUGGINGFACE_HUB_TOKEN が設定されていません。")
|
16 |
-
|
17 |
-
# キャッシュディレクトリの作成(書き込み可能な /tmp を利用)
|
18 |
-
cache_dir = "/tmp/hf_cache"
|
19 |
-
os.makedirs(cache_dir, exist_ok=True)
|
20 |
-
|
21 |
-
# pyannote モデルの読み込み
|
22 |
-
model = Model.from_pretrained("pyannote/embedding", use_auth_token=hf_token, cache_dir=cache_dir)
|
23 |
-
inference = Inference(model)
|
24 |
-
|
25 |
-
def cosine_similarity(vec1, vec2):
|
26 |
-
vec1 = vec1 / np.linalg.norm(vec1)
|
27 |
-
vec2 = vec2 / np.linalg.norm(vec2)
|
28 |
-
return np.dot(vec1, vec2)
|
29 |
-
|
30 |
-
def segment_audio(path, target_path='/tmp/setup_voice', seg_duration=1.0):
|
31 |
-
"""
|
32 |
-
音声を指定秒数ごとに分割する。
|
33 |
-
target_path に分割したファイルを保存し、元の音声の総長(ミリ秒)を返す。
|
34 |
-
"""
|
35 |
-
os.makedirs(target_path, exist_ok=True)
|
36 |
-
base_sound = AudioSegment.from_file(path)
|
37 |
-
duration_ms = len(base_sound)
|
38 |
-
seg_duration_ms = int(seg_duration * 1000)
|
39 |
-
|
40 |
-
for i, start in enumerate(range(0, duration_ms, seg_duration_ms)):
|
41 |
-
end = min(start + seg_duration_ms, duration_ms)
|
42 |
-
segment = base_sound[start:end]
|
43 |
-
segment.export(os.path.join(target_path, f'{i}.wav'), format="wav")
|
44 |
-
|
45 |
-
return target_path, duration_ms
|
46 |
-
|
47 |
-
def calculate_similarity(path1, path2):
|
48 |
-
embedding1 = inference(path1)
|
49 |
-
embedding2 = inference(path2)
|
50 |
-
return float(cosine_similarity(embedding1.data.flatten(), embedding2.data.flatten()))
|
51 |
-
|
52 |
-
def process_audio(reference_path, input_path, output_folder='/tmp/data/matched_segments', seg_duration=1.0, threshold=0.5):
|
53 |
-
"""
|
54 |
-
入力音声ファイルを seg_duration 秒ごとに分割し、各セグメントと参照音声の類似度を計算。
|
55 |
-
類似度が threshold を超えたセグメントを output_folder にコピーし、マッチした時間(ms)と
|
56 |
-
マッチしなかった時間(ms)を返す。
|
57 |
-
"""
|
58 |
-
os.makedirs(output_folder, exist_ok=True)
|
59 |
-
segmented_path, total_duration_ms = segment_audio(input_path, seg_duration=seg_duration)
|
60 |
-
|
61 |
-
matched_time_ms = 0
|
62 |
-
for file in sorted(os.listdir(segmented_path)):
|
63 |
-
segment_file = os.path.join(segmented_path, file)
|
64 |
-
similarity = calculate_similarity(segment_file, reference_path)
|
65 |
-
if similarity > threshold:
|
66 |
-
shutil.copy(segment_file, output_folder)
|
67 |
-
matched_time_ms += len(AudioSegment.from_file(segment_file))
|
68 |
-
|
69 |
-
unmatched_time_ms = total_duration_ms - matched_time_ms
|
70 |
-
return matched_time_ms, unmatched_time_ms
|
71 |
-
|
72 |
-
def generate_random_string(length):
|
73 |
-
letters = string.ascii_letters + string.digits
|
74 |
-
return ''.join(random.choice(letters) for i in range(length))
|
75 |
-
|
76 |
-
def generate_filename(random_length):
|
77 |
-
random_string = generate_random_string(random_length)
|
78 |
-
current_time = datetime.now().strftime("%Y%m%d%H%M%S")
|
79 |
-
filename = f"{current_time}_{random_string}.wav"
|
80 |
-
return filename
|
81 |
|
|
|
82 |
app = Flask(__name__)
|
83 |
|
84 |
# トップページ(テンプレート: index.html)
|
@@ -121,7 +48,7 @@ def upload_audio():
|
|
121 |
|
122 |
# 音声解析:参照音声とアップロードされた音声との類似度をセグメント毎に計算
|
123 |
# threshold の値は調整可能です(例: 0.1)
|
124 |
-
matched_time, unmatched_time = process_audio(reference_audio, audio_path, threshold=0.1)
|
125 |
total_time = matched_time + unmatched_time
|
126 |
rate = (matched_time / total_time) * 100 if total_time > 0 else 0
|
127 |
|
|
|
1 |
from flask import Flask, request, jsonify, render_template, send_from_directory
|
2 |
import base64
|
3 |
+
|
4 |
import os
|
5 |
import shutil
|
6 |
+
from process import AudioProcessor
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
7 |
|
8 |
+
process=AudioProcessor()
|
9 |
app = Flask(__name__)
|
10 |
|
11 |
# トップページ(テンプレート: index.html)
|
|
|
48 |
|
49 |
# 音声解析:参照音声とアップロードされた音声との類似度をセグメント毎に計算
|
50 |
# threshold の値は調整可能です(例: 0.1)
|
51 |
+
matched_time, unmatched_time = process.process_audio(reference_audio, audio_path, threshold=0.1)
|
52 |
total_time = matched_time + unmatched_time
|
53 |
rate = (matched_time / total_time) * 100 if total_time > 0 else 0
|
54 |
|
process.py
ADDED
@@ -0,0 +1,81 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
import os
|
3 |
+
import shutil
|
4 |
+
import numpy as np
|
5 |
+
import string
|
6 |
+
import random
|
7 |
+
from datetime import datetime
|
8 |
+
from pyannote.audio import Model, Inference
|
9 |
+
from pydub import AudioSegment
|
10 |
+
class AudioProcessor():
|
11 |
+
def __init__(self,cache_dir = "/tmp/hf_cache"):
|
12 |
+
hf_token = os.environ.get("HF")
|
13 |
+
if hf_token is None:
|
14 |
+
print('3')
|
15 |
+
raise ValueError("HUGGINGFACE_HUB_TOKEN が設定されていません。")
|
16 |
+
os.makedirs(cache_dir, exist_ok=True)
|
17 |
+
# pyannote モデルの読み込み
|
18 |
+
model = Model.from_pretrained("pyannote/embedding", use_auth_token=hf_token, cache_dir=cache_dir)
|
19 |
+
self.inference = Inference(model)
|
20 |
+
|
21 |
+
def cosine_similarity(self,vec1, vec2):
|
22 |
+
vec1 = vec1 / np.linalg.norm(vec1)
|
23 |
+
vec2 = vec2 / np.linalg.norm(vec2)
|
24 |
+
return np.dot(vec1, vec2)
|
25 |
+
|
26 |
+
def segment_audio(self, path, target_path='/tmp/setup_voice', seg_duration=1.0):
|
27 |
+
"""
|
28 |
+
音声を指定秒数ごとに分割し、短いセグメントには無音をパディングする。
|
29 |
+
"""
|
30 |
+
os.makedirs(target_path, exist_ok=True)
|
31 |
+
base_sound = AudioSegment.from_file(path)
|
32 |
+
duration_ms = len(base_sound)
|
33 |
+
seg_duration_ms = int(seg_duration * 1000)
|
34 |
+
|
35 |
+
for i, start in enumerate(range(0, duration_ms, seg_duration_ms)):
|
36 |
+
end = min(start + seg_duration_ms, duration_ms)
|
37 |
+
segment = base_sound[start:end]
|
38 |
+
# セグメントが指定長さに満たない場合、無音でパディングする
|
39 |
+
if len(segment) < seg_duration_ms:
|
40 |
+
silence = AudioSegment.silent(duration=(seg_duration_ms - len(segment)))
|
41 |
+
segment = segment + silence
|
42 |
+
|
43 |
+
segment.export(os.path.join(target_path, f'{i}.wav'), format="wav")
|
44 |
+
|
45 |
+
return target_path, duration_ms
|
46 |
+
|
47 |
+
|
48 |
+
def calculate_similarity(self,path1, path2):
|
49 |
+
embedding1 = self.inference(path1)
|
50 |
+
embedding2 = self.inference(path2)
|
51 |
+
return float(self.cosine_similarity(embedding1.data.flatten(), embedding2.data.flatten()))
|
52 |
+
|
53 |
+
def process_audio(self,reference_path, input_path, output_folder='/tmp/data/matched_segments', seg_duration=1.0, threshold=0.5):
|
54 |
+
"""
|
55 |
+
入力音声ファイルを seg_duration 秒ごとに分割し、各セグメントと参照音声の類似度を計算。
|
56 |
+
類似度が threshold を超えたセグメントを output_folder にコピーし、マッチした時間(ms)と
|
57 |
+
マッチしなかった時間(ms)を返す。
|
58 |
+
"""
|
59 |
+
os.makedirs(output_folder, exist_ok=True)
|
60 |
+
segmented_path, total_duration_ms = self.segment_audio(input_path, seg_duration=seg_duration)
|
61 |
+
|
62 |
+
matched_time_ms = 0
|
63 |
+
for file in sorted(os.listdir(segmented_path)):
|
64 |
+
segment_file = os.path.join(segmented_path, file)
|
65 |
+
similarity = self.calculate_similarity(segment_file, reference_path)
|
66 |
+
if similarity > threshold:
|
67 |
+
shutil.copy(segment_file, output_folder)
|
68 |
+
matched_time_ms += len(AudioSegment.from_file(segment_file))
|
69 |
+
|
70 |
+
unmatched_time_ms = total_duration_ms - matched_time_ms
|
71 |
+
return matched_time_ms, unmatched_time_ms
|
72 |
+
|
73 |
+
def generate_random_string(self,length):
|
74 |
+
letters = string.ascii_letters + string.digits
|
75 |
+
return ''.join(random.choice(letters) for i in range(length))
|
76 |
+
|
77 |
+
def generate_filename(self,random_length):
|
78 |
+
random_string = self.generate_random_string(random_length)
|
79 |
+
current_time = datetime.now().strftime("%Y%m%d%H%M%S")
|
80 |
+
filename = f"{current_time}_{random_string}.wav"
|
81 |
+
return filename
|