Update app.py

app.py

@@ -7,6 +7,9 @@ import os
 import re
 from pathlib import Path
 from pydub import AudioSegment
+import librosa
+import soundfile as sf
+import numpy as np
 
 def get_silence(duration_ms=1000):
     # Create silent audio segment with specified parameters
@@ -14,11 +17,9 @@ def get_silence(duration_ms=1000):
         duration=duration_ms,
         frame_rate=24000  # 24kHz sampling rate
     )
-
     # Set audio parameters
     silent_audio = silent_audio.set_channels(1)  # Mono
     silent_audio = silent_audio.set_sample_width(4)  # 32-bit (4 bytes per sample)
-
     with tempfile.NamedTemporaryFile(delete=False, suffix=".mp3") as tmp_file:
         # Export with specific bitrate and codec parameters
         silent_audio.export(
@@ -36,220 +37,183 @@ def get_silence(duration_ms=1000):
 
 # Get all available voices
 async def get_voices():
-    voices = await edge_tts.list_voices()
-    return {f"{v['ShortName']} - {v['Locale']} ({v['Gender']})": v['ShortName'] for v in voices}
+    try:
+        voices = await edge_tts.list_voices()
+        return {f"{v['ShortName']} - {v['Locale']} ({v['Gender']})": v['ShortName'] for v in voices}
+    except Exception as e:
+        print(f"Error listing voices: {e}")
+        return {}
 
-async def generate_audio_with_voice_prefix(text_segment, default_voice, rate, pitch):
-    """Generates audio for a text segment, handling voice prefixes."""
+async def generate_audio_with_voice_prefix(text_segment, default_voice, rate, pitch, target_duration_ms=None, speed_adjustment_factor=1.0):
+    """Generates audio for a text segment, handling voice prefixes and adjusting rate for duration."""
     current_voice_full = default_voice
     current_voice_short = current_voice_full.split(" - ")[0] if current_voice_full else ""
     current_rate = rate
     current_pitch = pitch
     processed_text = text_segment.strip()
-    voice1_full = "en-AU-WilliamNeural - en-AU (Male)"
-    voice1_short = voice1_full.split(" - ")[0]
-    voice1F_full ="en-GB-SoniaNeural - en-GB (Female)"
-    voice1F_short = voice1F_full.split(" - ")[0]
-    voice2_full = "en-GB-RyanNeural - en-GB (Male)"
-    voice2_short = voice2_full.split(" - ")[0]
-    voice2F_full = "en-US-JennyNeural - en-US (Female)"
-    voice2F_short = voice2F_full.split(" - ")[0]
-    voice3_full ="en-US-BrianMultilingualNeural - en-US (Male)"  #good for reading
-    voice3_short = voice3_full.split(" - ")[0]
-    voice3F_full = "en-HK-YanNeural - en-HK (Female)"
-    voice3F_short = voice3F_full.split(" - ")[0]
-    voice4_full = "en-GB-ThomasNeural - en-GB (Male)"
-    voice4_short = voice4_full.split(" - ")[0]
-    voice4F_full ="en-US-EmmaNeural - en-US (Female)"
-    voice4F_short = voice4F_full.split(" - ")[0]
-    voice5_full = "en-GB-RyanNeural - en-GB (Male)" #Old Man
-    voice5_short = voice5_full.split(" - ")[0]
-    voice6_full = "en-GB-MaisieNeural - en-GB (Female)"  #Child
-    voice6_short = voice6_full.split(" - ")[0]
-    voice7_full = "vi-VN-HoaiMyNeural - vi-VN (Female)"  #Vietnamese
-    voice7_short = voice7_full.split(" - ")[0]
-    voice8_full = "vi-VN-NamMinhNeural - vi-VN (Male)"  #Vietnamese
-    voice8_short = voice8_full.split(" - ")[0]
-    voice9F_full = "de-DE-SeraphinaMultilingualNeural - de-DE (Female)"  #Vietnamese
-    voice9F_short = voice7_full.split(" - ")[0]
-    voice9_full = "ko-KR-HyunsuMultilingualNeural - ko-KR (Male)"  #Vietnamese
-    voice9_short = voice8_full.split(" - ")[0]
-    detect=0
-    if processed_text.startswith("1F"):
-        current_voice_short = voice1F_short
-        current_pitch = 25
-        detect=1
-        #processed_text = processed_text[2:].strip()
-    elif processed_text.startswith("2F"):
-        current_voice_short = voice2F_short
-        #processed_text = processed_text[2:].strip()
-        detect=1
-    elif processed_text.startswith("3F"):
-        current_voice_short = voice3F_short
-        #processed_text = processed_text[2:].strip()
-        detect=1
-    elif processed_text.startswith("4F"):
-        current_voice_short = voice4F_short
-        #processed_text = processed_text[2:].strip()
-        detect=1
-    elif processed_text.startswith("1M"):
-        current_voice_short = voice1_short
-        #processed_text = processed_text[2:].strip()
-        detect=1
-    elif processed_text.startswith("2M"):
-        current_voice_short = voice2_short
-        #processed_text = processed_text[2:].strip()
-        detect=1
-    elif processed_text.startswith("3M"):
-        current_voice_short = voice3_short
-        #processed_text = processed_text[2:].strip()
-        detect=1
-    elif processed_text.startswith("4M"):
-        current_voice_short = voice4_short
-        #processed_text = processed_text[2:].strip()
-        detect=1
-    elif processed_text.startswith("1O"):  # Old man voice
-        current_voice_short = voice5_short
-        current_pitch = -20
-        current_rate = -10
-        #processed_text = processed_text[2:].strip()
-        detect=1
-    elif processed_text.startswith("1C"):  #Child voice
-        current_voice_short = voice6_short
-        #processed_text = processed_text[2:].strip()
-        detect=1
-    elif processed_text.startswith("1V"):  #Female VN
-        current_voice_short = voice7_short
-        #processed_text = processed_text[2:].strip()
-        detect=1
-    elif processed_text.startswith("2V"):
-        current_voice_short = voice8_short
-        #processed_text = processed_text[2:].strip()
-        detect=1
-    elif processed_text.startswith("3V"):  #Female VN
-        current_voice_short = voice9F_short
-        current_pitch = 25
-        #processed_text = processed_text[2:].strip()
-        detect=1
-    elif processed_text.startswith("4V"):
-        current_voice_short = voice9_short
-        current_pitch = -20
-        #processed_text = processed_text[2:].strip()
-        detect=1
-    #Looking for number following prefix, which are pitch values.
-    #match = re.search(r'[A-Za-z]\d+', part)  # Look for a letter followed by one or more digits
-    match = re.search(r'[A-Za-z]+\-?\d+', processed_text)  # Look for a letter(s) followed by an optional '-' and digits
+    print(f"Processing this  text segment: {processed_text}") # Debug
+    voice_map = {
+        "1F": "en-GB-SoniaNeural",
+        "2M": "en-GB-RyanNeural",
+        "3M": "en-US-BrianMultilingualNeural",
+        "2F": "en-US-JennyNeural",
+        "1M": "en-AU-WilliamNeural",
+        "3F": "en-HK-YanNeural",
+        "4M": "en-GB-ThomasNeural",
+        "4F": "en-US-EmmaNeural",
+        "1O": "en-GB-RyanNeural",  # Old Man
+        "1C": "en-GB-MaisieNeural",  # Child
+        "1V": "vi-VN-HoaiMyNeural",  # Vietnamese (Female)
+        "2V": "vi-VN-NamMinhNeural",  # Vietnamese (Male)
+        "3V": "vi-VN-HoaiMyNeural",  # Vietnamese (Female)
+        "4V": "vi-VN-NamMinhNeural",  # Vietnamese (Male)
+    }
+    detect = 0
+    for prefix, voice_short in voice_map.items():
+        if processed_text.startswith(prefix):
+            current_voice_short = voice_short
+            if prefix in ["1F", "3F", "1V", "3V"]:
+                current_pitch = 25
+            elif prefix in ["1O", "4V"]:
+                current_pitch = -20
+                current_rate = -10
+            detect = 1
+            processed_text = processed_text[len(prefix):].strip()
+            break
+    match = re.search(r'([A-Za-z]+)-?(\d+)', processed_text)
     if match:
-        # Extract the prefix (e.g., '2F') and number (e.g., '-20')
-        prefix = ''.join([ch for ch in match.group() if ch.isalpha()])  # Extract letters (prefix)
-        number = int(''.join([ch for ch in match.group() if ch.isdigit() or ch == '-']))  # Extract digits (number)
-        current_pitch += number
-        # Step 2: Remove the found number from the string
-        new_text = re.sub(r'[A-Za-z]+\-?\d+', '', processed_text, count=1).strip()  # Remove prefix and number (e.g., '2F-20')
-        #processed_text = new_text[2:]  #cut out the prefix like 1F, 3M etc
-        processed_text = new_text[len(prefix):]  # Dynamically remove the prefix part
-    else:
-        if detect:
-            processed_text = processed_text[2:]
+        prefix_pitch = match.group(1)
+        number = int(match.group(2))
+        if prefix_pitch in voice_map:
+            current_pitch += number
+            processed_text = re.sub(r'[A-Za-z]+-?\d+', '', processed_text, count=1).strip()
+        elif detect:
+            processed_text = processed_text.lstrip('-0123456789').strip() # Remove potential leftover numbers
+    elif detect:
+        processed_text = processed_text[2:].strip()
     if processed_text:
         rate_str = f"{current_rate:+d}%"
         pitch_str = f"{current_pitch:+d}Hz"
-        communicate = edge_tts.Communicate(processed_text, current_voice_short, rate=rate_str, pitch=pitch_str)
-        with tempfile.NamedTemporaryFile(delete=False, suffix=".mp3") as tmp_file:
-            audio_path = tmp_file.name
-            await communicate.save(audio_path)
-        return audio_path
+        try:
+            communicate = edge_tts.Communicate(processed_text, current_voice_short, rate=rate_str, pitch=pitch_str)
+            with tempfile.NamedTemporaryFile(delete=False, suffix=".mp3") as tmp_file:
+                audio_path = tmp_file.name
+                await communicate.save(audio_path)
+            if target_duration_ms is not None and os.path.exists(audio_path):
+                audio = AudioSegment.from_mp3(audio_path)
+                audio_duration_ms = len(audio)
+                #print(f"Generated audio duration: {audio_duration_ms}ms, Target duration: {target_duration_ms}ms") # Debug
+                if audio_duration_ms > target_duration_ms and target_duration_ms > 0:
+                    speed_factor = (audio_duration_ms / target_duration_ms) * speed_adjustment_factor
+                    #print(f"Speed factor (after user adjustment): {speed_factor}") # Debug
+                    if speed_factor > 0:
+                        if speed_factor < 1.0:
+                            speed_factor = 1.0
+                        y, sr = librosa.load(audio_path, sr=None)
+                        y_stretched = librosa.effects.time_stretch(y, rate=speed_factor)
+                        sf.write(audio_path, y_stretched, sr)
+                else:
+                    print("Generated audio is not longer than target duration, no speed adjustment.") # Debug
+            return audio_path
+        except Exception as e:
+            print(f"Edge TTS error processing '{processed_text}': {e}")
+            return None
     return None
 
-async def process_transcript_line(line, default_voice, rate, pitch):
-    """Processes a single transcript line with HH:MM:SS.milliseconds timestamp and quoted text segments."""
-    match = re.match(r'(\d{2}):(\d{2}):(\d{2})\.(\d{3})\s+(.*)', line)
+async def process_transcript_line(line, default_voice, rate, pitch, speed_adjustment_factor):
+    """Processes a single transcript line with HH:MM:SS,milliseconds - HH:MM:SS,milliseconds timestamp."""
+    match = re.match(r'(\d{2}):(\d{2}):(\d{2}),(\d{3})\s+-\s+(\d{2}):(\d{2}):(\d{2}),(\d{3})\s+(.*)', line)
     if match:
-        hours, minutes, seconds, milliseconds, text_parts = match.groups()
+        start_h, start_m, start_s, start_ms, end_h, end_m, end_s, end_ms, text_parts = match.groups()
         start_time_ms = (
-            int(hours) * 3600000 +
-            int(minutes) * 60000 +
-            int(seconds) * 1000 +
-            int(milliseconds)
+            int(start_h) * 3600000 +
+            int(start_m) * 60000 +
+            int(start_s) * 1000 +
+            int(start_ms)
+        )
+        end_time_ms = (
+            int(end_h) * 3600000 +
+            int(end_m) * 60000 +
+            int(end_s) * 1000 +
+            int(end_ms)
         )
+        duration_ms = end_time_ms - start_time_ms
         audio_segments = []
-        split_parts = re.split(r'(")', text_parts)  # Split by quote marks, keeping the quotes
-
+        split_parts = re.split(r'[“”"]', text_parts)
         process_next = False
         for part in split_parts:
             if part == '"':
                 process_next = not process_next
                 continue
             if process_next and part.strip():
-                audio_path = await generate_audio_with_voice_prefix(part, default_voice, rate, pitch)
+                audio_path = await generate_audio_with_voice_prefix(part, default_voice, rate, pitch, duration_ms, speed_adjustment_factor)
                 if audio_path:
                     audio_segments.append(audio_path)
             elif not process_next and part.strip():
-                audio_path = await generate_audio_with_voice_prefix(part, default_voice, rate, pitch) # Process unquoted text with default voice
+                audio_path = await generate_audio_with_voice_prefix(part, default_voice, rate, pitch, duration_ms, speed_adjustment_factor)
                 if audio_path:
                     audio_segments.append(audio_path)
+        return start_time_ms, audio_segments, duration_ms
+    return None, None, None
 
-        return start_time_ms, audio_segments
-    return None, None
-
-async def transcript_to_speech(transcript_text, voice, rate, pitch):
+async def transcript_to_speech(transcript_text, voice, rate, pitch, speed_adjustment_factor):
     if not transcript_text.strip():
         return None, gr.Warning("Please enter transcript text.")
     if not voice:
         return None, gr.Warning("Please select a voice.")
-
     lines = transcript_text.strip().split('\n')
     timed_audio_segments = []
     max_end_time_ms = 0
-
     for line in lines:
-        start_time, audio_paths = await process_transcript_line(line, voice, rate, pitch)
+        start_time, audio_paths, duration = await process_transcript_line(line, voice, rate, pitch, speed_adjustment_factor)
         if start_time is not None and audio_paths:
             combined_line_audio = AudioSegment.empty()
+            current_time_ms = start_time
+            segment_duration = duration / len(audio_paths) if audio_paths else 0
             for path in audio_paths:
-                try:
-                    audio = AudioSegment.from_mp3(path)
-                    combined_line_audio += audio
-                    os.remove(path)
-                except FileNotFoundError:
-                    print(f"Warning: Audio file not found: {path}")
-
+                if path:  # Only process if audio_path is not None (meaning TTS was successful)
+                    try:
+                        audio = AudioSegment.from_mp3(path)
+                        combined_line_audio += audio
+                        os.remove(path)
+                    except FileNotFoundError:
+                        print(f"Warning: Audio file not found: {path}")
             if combined_line_audio:
                 timed_audio_segments.append({'start': start_time, 'audio': combined_line_audio})
                 max_end_time_ms = max(max_end_time_ms, start_time + len(combined_line_audio))
         elif audio_paths:
             for path in audio_paths:
-                try:
-                    os.remove(path)
-                except FileNotFoundError:
-                    pass # Clean up even if no timestamp
-
+                if path:
+                    try:
+                        os.remove(path)
+                    except FileNotFoundError:
+                        pass # Clean up even if no timestamp
     if not timed_audio_segments:
         return None, "No processable audio segments found."
-
     final_audio = AudioSegment.silent(duration=max_end_time_ms, frame_rate=24000)
     for segment in timed_audio_segments:
         final_audio = final_audio.overlay(segment['audio'], position=segment['start'])
-
     combined_audio_path = tempfile.mktemp(suffix=".mp3")
     final_audio.export(combined_audio_path, format="mp3")
     return combined_audio_path, None
 
 @spaces.GPU
-def tts_interface(transcript, voice, rate, pitch):
-    audio, warning = asyncio.run(transcript_to_speech(transcript, voice, rate, pitch))
+def tts_interface(transcript, voice, rate, pitch, speed_adjustment_factor):
+    audio, warning = asyncio.run(transcript_to_speech(transcript, voice, rate, pitch, speed_adjustment_factor))
     return audio, warning
 
 async def create_demo():
     voices = await get_voices()
     default_voice = "en-US-AndrewMultilingualNeural - en-US (Male)"
     description = """
-    Process timestamped text (HH:MM:SS.milliseconds) with voice changes within quotes.
-    Format: `HH:MM:SS.milliseconds "VoicePrefix Text" more text "AnotherVoicePrefix More Text"`
+    Process timestamped text (HH:MM:SS,milliseconds - HH:MM:SS,milliseconds) with voice changes within quotes.
+    The duration specified in the timestamp will be used to adjust the speech rate so the generated audio fits within that time.
+    You can control the intensity of the speed adjustment using the "Speed Adjustment Factor" slider.
+    Format: `HH:MM:SS,milliseconds - HH:MM:SS,milliseconds "VoicePrefix Text" more text "AnotherVoicePrefix More Text"`
     Example:
     ```
-    00:00:00.000 "This is the default voice." more default. "1F Now a female voice." and back to default.
-    00:00:05.000 "1C Yes," said the child, "it is fun!"
+    00:00:00,000 - 00:00:05,000 "This is the default voice." more default. "1F Now a female voice." and back to default.
+    00:00:05,500 - 00:00:10,250 "1C Yes," said the child, "it is fun!"
     ```
     ***************************************************************************************************
     1M = en-AU-WilliamNeural - en-AU (Male)
@@ -271,16 +235,17 @@ async def create_demo():
     demo = gr.Interface(
         fn=tts_interface,
         inputs=[
-            gr.Textbox(label="Timestamped Text with Voice Changes", lines=10, placeholder='00:00:00.000 "Text" more text "1F Different Voice"'),
+            gr.Textbox(label="Timestamped Text with Voice Changes and Duration", lines=10, placeholder='00:00:00,000 - 00:00:05,000 "Text" more text "1F Different Voice"'),
             gr.Dropdown(choices=[""] + list(voices.keys()), label="Select Default Voice", value=default_voice),
             gr.Slider(minimum=-50, maximum=50, value=0, label="Speech Rate Adjustment (%)", step=1),
-            gr.Slider(minimum=-50, maximum=50, value=0, label="Pitch Adjustment (Hz)", step=1)
+            gr.Slider(minimum=-50, maximum=50, value=0, label="Pitch Adjustment (Hz)", step=1),
+            gr.Slider(minimum=0.5, maximum=1.5, value=1.0, step=0.05, label="Speed Adjustment Factor")
         ],
         outputs=[
             gr.Audio(label="Generated Audio", type="filepath"),
             gr.Markdown(label="Warning", visible=False)
         ],
-        title="TTS with HH:MM:SS.milliseconds and In-Quote Voice Switching",
+        title="TTS with Duration-Aware Speed Adjustment and In-Quote Voice Switching",
         description=description,
         analytics_enabled=False,
         allow_flagging=False
@@ -289,4 +254,4 @@ async def create_demo():
 
 if __name__ == "__main__":
     demo = asyncio.run(create_demo())
-    demo.launch()
+    demo.launch()