Update app.py
Browse files
app.py
CHANGED
@@ -11,32 +11,35 @@ import librosa
|
|
11 |
import soundfile as sf
|
12 |
import numpy as np
|
13 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
def get_silence(duration_ms=1000):
|
15 |
-
|
16 |
-
|
17 |
duration=duration_ms,
|
18 |
-
frame_rate=24000
|
|
|
|
|
19 |
)
|
20 |
-
# Set audio parameters
|
21 |
-
silent_audio = silent_audio.set_channels(1) # Mono
|
22 |
-
silent_audio = silent_audio.set_sample_width(4) # 32-bit (4 bytes per sample)
|
23 |
-
with tempfile.NamedTemporaryFile(delete=False, suffix=".mp3") as tmp_file:
|
24 |
-
# Export with specific bitrate and codec parameters
|
25 |
-
silent_audio.export(
|
26 |
-
tmp_file.name,
|
27 |
-
format="mp3",
|
28 |
-
bitrate="48k",
|
29 |
-
parameters=[
|
30 |
-
"-ac", "1", # Mono
|
31 |
-
"-ar", "24000", # Sample rate
|
32 |
-
"-sample_fmt", "s32", # 32-bit samples
|
33 |
-
"-codec:a", "libmp3lame" # MP3 codec
|
34 |
-
]
|
35 |
-
)
|
36 |
-
return tmp_file.name
|
37 |
|
38 |
-
# Get all available voices
|
39 |
async def get_voices():
|
|
|
40 |
try:
|
41 |
voices = await edge_tts.list_voices()
|
42 |
return {f"{v['ShortName']} - {v['Locale']} ({v['Gender']})": v['ShortName'] for v in voices}
|
@@ -46,30 +49,12 @@ async def get_voices():
|
|
46 |
|
47 |
async def generate_audio_with_voice_prefix(text_segment, default_voice, rate, pitch, target_duration_ms=None, speed_adjustment_factor=1.0):
|
48 |
"""Generates audio for a text segment, handling voice prefixes and adjusting rate for duration."""
|
49 |
-
|
50 |
-
current_voice_short =
|
51 |
current_rate = rate
|
52 |
current_pitch = pitch
|
53 |
-
|
54 |
-
|
55 |
-
voice_map = {
|
56 |
-
"1F": "en-GB-SoniaNeural",
|
57 |
-
"2M": "en-GB-RyanNeural",
|
58 |
-
"3M": "en-US-BrianMultilingualNeural",
|
59 |
-
"2F": "en-US-JennyNeural",
|
60 |
-
"1M": "en-AU-WilliamNeural",
|
61 |
-
"3F": "en-HK-YanNeural",
|
62 |
-
"4M": "en-GB-ThomasNeural",
|
63 |
-
"4F": "en-US-EmmaNeural",
|
64 |
-
"1O": "en-GB-RyanNeural", # Old Man
|
65 |
-
"1C": "en-GB-MaisieNeural", # Child
|
66 |
-
"1V": "vi-VN-HoaiMyNeural", # Vietnamese (Female)
|
67 |
-
"2V": "vi-VN-NamMinhNeural", # Vietnamese (Male)
|
68 |
-
"3V": "vi-VN-HoaiMyNeural", # Vietnamese (Female)
|
69 |
-
"4V": "vi-VN-NamMinhNeural", # Vietnamese (Male)
|
70 |
-
}
|
71 |
-
detect = 0
|
72 |
-
for prefix, voice_short in voice_map.items():
|
73 |
if processed_text.startswith(prefix):
|
74 |
current_voice_short = voice_short
|
75 |
if prefix in ["1F", "3F", "1V", "3V"]:
|
@@ -77,20 +62,17 @@ async def generate_audio_with_voice_prefix(text_segment, default_voice, rate, pi
|
|
77 |
elif prefix in ["1O", "4V"]:
|
78 |
current_pitch = -20
|
79 |
current_rate = -10
|
80 |
-
detect = 1
|
81 |
processed_text = processed_text[len(prefix):].strip()
|
82 |
break
|
|
|
83 |
match = re.search(r'([A-Za-z]+)-?(\d+)', processed_text)
|
84 |
-
if match:
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
processed_text = processed_text.lstrip('-0123456789').strip() # Remove potential leftover numbers
|
92 |
-
elif detect:
|
93 |
-
processed_text = processed_text[2:].strip()
|
94 |
if processed_text:
|
95 |
rate_str = f"{current_rate:+d}%"
|
96 |
pitch_str = f"{current_pitch:+d}Hz"
|
@@ -99,110 +81,109 @@ async def generate_audio_with_voice_prefix(text_segment, default_voice, rate, pi
|
|
99 |
with tempfile.NamedTemporaryFile(delete=False, suffix=".mp3") as tmp_file:
|
100 |
audio_path = tmp_file.name
|
101 |
await communicate.save(audio_path)
|
102 |
-
|
103 |
-
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
y_stretched = librosa.effects.time_stretch(y, rate=speed_factor)
|
114 |
-
sf.write(audio_path, y_stretched, sr)
|
115 |
-
else:
|
116 |
-
print("Generated audio is not longer than target duration, no speed adjustment.") # Debug
|
117 |
-
return audio_path
|
118 |
except Exception as e:
|
119 |
print(f"Edge TTS error processing '{processed_text}': {e}")
|
120 |
return None
|
121 |
return None
|
122 |
|
123 |
async def process_transcript_line(line, default_voice, rate, pitch, speed_adjustment_factor):
|
124 |
-
"""Processes a single transcript line with
|
125 |
-
match = re.match(r'(\d{2}
|
126 |
if match:
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
-
int(
|
133 |
-
|
134 |
-
|
135 |
-
|
136 |
-
int(end_m) * 60000 +
|
137 |
-
int(end_s) * 1000 +
|
138 |
-
int(end_ms)
|
139 |
-
)
|
140 |
duration_ms = end_time_ms - start_time_ms
|
|
|
141 |
audio_segments = []
|
142 |
-
|
143 |
-
|
144 |
-
for part in
|
145 |
if part == '"':
|
146 |
-
|
147 |
continue
|
148 |
-
if
|
149 |
-
audio_path = await generate_audio_with_voice_prefix(part, default_voice, rate, pitch, duration_ms, speed_adjustment_factor)
|
150 |
-
if audio_path:
|
151 |
-
audio_segments.append(audio_path)
|
152 |
-
elif not process_next and part.strip():
|
153 |
-
audio_path = await generate_audio_with_voice_prefix(part, default_voice, rate, pitch, duration_ms, speed_adjustment_factor)
|
154 |
if audio_path:
|
155 |
audio_segments.append(audio_path)
|
156 |
return start_time_ms, audio_segments, duration_ms
|
157 |
return None, None, None
|
158 |
|
159 |
async def transcript_to_speech(transcript_text, voice, rate, pitch, speed_adjustment_factor):
|
|
|
160 |
if not transcript_text.strip():
|
161 |
return None, gr.Warning("Please enter transcript text.")
|
162 |
if not voice:
|
163 |
return None, gr.Warning("Please select a voice.")
|
|
|
164 |
lines = transcript_text.strip().split('\n')
|
165 |
timed_audio_segments = []
|
166 |
max_end_time_ms = 0
|
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 |
@spaces.GPU
|
201 |
def tts_interface(transcript, voice, rate, pitch, speed_adjustment_factor):
|
|
|
202 |
audio, warning = asyncio.run(transcript_to_speech(transcript, voice, rate, pitch, speed_adjustment_factor))
|
203 |
return audio, warning
|
204 |
|
205 |
async def create_demo():
|
|
|
206 |
voices = await get_voices()
|
207 |
default_voice = "en-US-AndrewMultilingualNeural - en-US (Male)"
|
208 |
description = """
|
|
|
11 |
import soundfile as sf
|
12 |
import numpy as np
|
13 |
|
14 |
+
# Global constant for voice mapping
|
15 |
+
VOICE_MAP = {
|
16 |
+
"1F": "en-GB-SoniaNeural",
|
17 |
+
"2M": "en-GB-RyanNeural",
|
18 |
+
"3M": "en-US-BrianMultilingualNeural",
|
19 |
+
"2F": "en-US-JennyNeural",
|
20 |
+
"1M": "en-AU-WilliamNeural",
|
21 |
+
"3F": "en-HK-YanNeural",
|
22 |
+
"4M": "en-GB-ThomasNeural",
|
23 |
+
"4F": "en-US-EmmaNeural",
|
24 |
+
"1O": "en-GB-RyanNeural", # Old Man
|
25 |
+
"1C": "en-GB-MaisieNeural", # Child
|
26 |
+
"1V": "vi-VN-HoaiMyNeural", # Vietnamese (Female)
|
27 |
+
"2V": "vi-VN-NamMinhNeural", # Vietnamese (Male)
|
28 |
+
"3V": "vi-VN-HoaiMyNeural", # Vietnamese (Female)
|
29 |
+
"4V": "vi-VN-NamMinhNeural", # Vietnamese (Male)
|
30 |
+
}
|
31 |
+
|
32 |
def get_silence(duration_ms=1000):
|
33 |
+
"""Creates a silent AudioSegment."""
|
34 |
+
return AudioSegment.silent(
|
35 |
duration=duration_ms,
|
36 |
+
frame_rate=24000,
|
37 |
+
sample_width=4,
|
38 |
+
channels=1
|
39 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
40 |
|
|
|
41 |
async def get_voices():
|
42 |
+
"""Lists available Edge TTS voices."""
|
43 |
try:
|
44 |
voices = await edge_tts.list_voices()
|
45 |
return {f"{v['ShortName']} - {v['Locale']} ({v['Gender']})": v['ShortName'] for v in voices}
|
|
|
49 |
|
50 |
async def generate_audio_with_voice_prefix(text_segment, default_voice, rate, pitch, target_duration_ms=None, speed_adjustment_factor=1.0):
|
51 |
"""Generates audio for a text segment, handling voice prefixes and adjusting rate for duration."""
|
52 |
+
processed_text = text_segment.strip()
|
53 |
+
current_voice_short = default_voice.split(" - ")[0] if default_voice else ""
|
54 |
current_rate = rate
|
55 |
current_pitch = pitch
|
56 |
+
|
57 |
+
for prefix, voice_short in VOICE_MAP.items():
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
58 |
if processed_text.startswith(prefix):
|
59 |
current_voice_short = voice_short
|
60 |
if prefix in ["1F", "3F", "1V", "3V"]:
|
|
|
62 |
elif prefix in ["1O", "4V"]:
|
63 |
current_pitch = -20
|
64 |
current_rate = -10
|
|
|
65 |
processed_text = processed_text[len(prefix):].strip()
|
66 |
break
|
67 |
+
|
68 |
match = re.search(r'([A-Za-z]+)-?(\d+)', processed_text)
|
69 |
+
if match and match.group(1) in VOICE_MAP:
|
70 |
+
pitch_adjustment = int(match.group(2))
|
71 |
+
current_pitch += pitch_adjustment
|
72 |
+
processed_text = re.sub(r'[A-Za-z]+-?\d+', '', processed_text, count=1).strip()
|
73 |
+
elif any(processed_text.startswith(prefix) for prefix in VOICE_MAP): # Handle leftover prefixes
|
74 |
+
processed_text = re.sub(r'^[A-Za-z]{1,2}', '', processed_text).lstrip('-').strip()
|
75 |
+
|
|
|
|
|
|
|
76 |
if processed_text:
|
77 |
rate_str = f"{current_rate:+d}%"
|
78 |
pitch_str = f"{current_pitch:+d}Hz"
|
|
|
81 |
with tempfile.NamedTemporaryFile(delete=False, suffix=".mp3") as tmp_file:
|
82 |
audio_path = tmp_file.name
|
83 |
await communicate.save(audio_path)
|
84 |
+
|
85 |
+
if target_duration_ms is not None and os.path.exists(audio_path) and target_duration_ms > 0:
|
86 |
+
audio = AudioSegment.from_mp3(audio_path)
|
87 |
+
audio_duration_ms = len(audio)
|
88 |
+
if audio_duration_ms > target_duration_ms:
|
89 |
+
speed_factor = (audio_duration_ms / target_duration_ms) * speed_adjustment_factor
|
90 |
+
if speed_factor > 0 and speed_factor >= 1.0:
|
91 |
+
y, sr = librosa.load(audio_path, sr=None)
|
92 |
+
y_stretched = librosa.effects.time_stretch(y, rate=speed_factor)
|
93 |
+
sf.write(audio_path, y_stretched, sr)
|
94 |
+
return audio_path
|
|
|
|
|
|
|
|
|
|
|
95 |
except Exception as e:
|
96 |
print(f"Edge TTS error processing '{processed_text}': {e}")
|
97 |
return None
|
98 |
return None
|
99 |
|
100 |
async def process_transcript_line(line, default_voice, rate, pitch, speed_adjustment_factor):
|
101 |
+
"""Processes a single transcript line with timestamp and potential voice changes."""
|
102 |
+
match = re.match(r'(\d{2}:\d{2}:\d{2},\d{3})\s+-\s+(\d{2}:\d{2}:\d{2},\d{3})\s+(.*)', line)
|
103 |
if match:
|
104 |
+
start_time_str, end_time_str, text_parts = match.groups()
|
105 |
+
|
106 |
+
def time_str_to_ms(time_str):
|
107 |
+
h, m, s_ms = time_str.split(':')
|
108 |
+
s, ms = s_ms.split(',')
|
109 |
+
return int(h) * 3600000 + int(m) * 60000 + int(s) * 1000 + int(ms)
|
110 |
+
|
111 |
+
start_time_ms = time_str_to_ms(start_time_str)
|
112 |
+
end_time_ms = time_str_to_ms(end_time_str)
|
|
|
|
|
|
|
|
|
113 |
duration_ms = end_time_ms - start_time_ms
|
114 |
+
|
115 |
audio_segments = []
|
116 |
+
parts = re.split(r'([“”"])', text_parts)
|
117 |
+
in_quote = False
|
118 |
+
for part in parts:
|
119 |
if part == '"':
|
120 |
+
in_quote = not in_quote
|
121 |
continue
|
122 |
+
if part.strip():
|
123 |
+
audio_path = await generate_audio_with_voice_prefix(part, default_voice, rate, pitch, duration_ms, speed_adjustment_factor if in_quote else 1.0)
|
|
|
|
|
|
|
|
|
124 |
if audio_path:
|
125 |
audio_segments.append(audio_path)
|
126 |
return start_time_ms, audio_segments, duration_ms
|
127 |
return None, None, None
|
128 |
|
129 |
async def transcript_to_speech(transcript_text, voice, rate, pitch, speed_adjustment_factor):
|
130 |
+
"""Converts a timestamped transcript with voice changes to a single audio file."""
|
131 |
if not transcript_text.strip():
|
132 |
return None, gr.Warning("Please enter transcript text.")
|
133 |
if not voice:
|
134 |
return None, gr.Warning("Please select a voice.")
|
135 |
+
|
136 |
lines = transcript_text.strip().split('\n')
|
137 |
timed_audio_segments = []
|
138 |
max_end_time_ms = 0
|
139 |
+
|
140 |
+
with tempfile.TemporaryDirectory() as tmpdir:
|
141 |
+
for line in lines:
|
142 |
+
start_time, audio_paths, duration = await process_transcript_line(line, voice, rate, pitch, speed_adjustment_factor)
|
143 |
+
if start_time is not None and audio_paths:
|
144 |
+
combined_line_audio = AudioSegment.empty()
|
145 |
+
for path in audio_paths:
|
146 |
+
if path and os.path.exists(path):
|
147 |
+
try:
|
148 |
+
audio = AudioSegment.from_mp3(path)
|
149 |
+
combined_line_audio += audio
|
150 |
+
except FileNotFoundError:
|
151 |
+
print(f"Warning: Audio file not found: {path}")
|
152 |
+
finally:
|
153 |
+
try:
|
154 |
+
os.remove(path)
|
155 |
+
except OSError:
|
156 |
+
print(f"Warning: Could not remove temporary file: {path}")
|
157 |
+
if combined_line_audio:
|
158 |
+
timed_audio_segments.append({'start': start_time, 'audio': combined_line_audio})
|
159 |
+
max_end_time_ms = max(max_end_time_ms, start_time + len(combined_line_audio))
|
160 |
+
elif audio_paths:
|
161 |
+
for path in audio_paths:
|
162 |
+
if path:
|
163 |
+
try:
|
164 |
+
os.remove(path)
|
165 |
+
except FileNotFoundError:
|
166 |
+
pass # Clean up even if no timestamp
|
167 |
+
|
168 |
+
if not timed_audio_segments:
|
169 |
+
return None, "No processable audio segments found."
|
170 |
+
|
171 |
+
final_audio = AudioSegment.silent(duration=max_end_time_ms, frame_rate=24000)
|
172 |
+
for segment in timed_audio_segments:
|
173 |
+
final_audio = final_audio.overlay(segment['audio'], position=segment['start'])
|
174 |
+
|
175 |
+
combined_audio_path = Path(tmpdir) / "combined_audio.mp3"
|
176 |
+
final_audio.export(str(combined_audio_path), format="mp3")
|
177 |
+
return str(combined_audio_path), None
|
178 |
|
179 |
@spaces.GPU
|
180 |
def tts_interface(transcript, voice, rate, pitch, speed_adjustment_factor):
|
181 |
+
"""Gradio interface function for TTS."""
|
182 |
audio, warning = asyncio.run(transcript_to_speech(transcript, voice, rate, pitch, speed_adjustment_factor))
|
183 |
return audio, warning
|
184 |
|
185 |
async def create_demo():
|
186 |
+
"""Creates the Gradio demo interface."""
|
187 |
voices = await get_voices()
|
188 |
default_voice = "en-US-AndrewMultilingualNeural - en-US (Male)"
|
189 |
description = """
|