app.py

import spaces
import gradio as gr
import edge_tts
import asyncio
import tempfile
import os
import re
from pathlib import Path
from pydub import AudioSegment

def get_silence(duration_ms=1000):
    # Create silent audio segment with specified parameters
    silent_audio = AudioSegment.silent(
        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(
            tmp_file.name,
            format="mp3",
            bitrate="48k",
            parameters=[
                "-ac", "1",  # Mono
                "-ar", "24000",  # Sample rate
                "-sample_fmt", "s32",  # 32-bit samples
                "-codec:a", "libmp3lame"  # MP3 codec
            ]
        )
        return tmp_file.name

# 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}

async def generate_audio_with_voice_prefix(text_segment, default_voice, rate, pitch):
    """Generates audio for a text segment, handling voice prefixes."""
    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
    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:]
    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
    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) # Modified timestamp regex
    if match:
        hours, minutes, seconds, milliseconds, text_parts = match.groups()
        start_time_ms = (
            int(hours) * 3600000 +
            int(minutes) * 60000 +
            int(seconds) * 1000 +
            int(milliseconds)
        )
        audio_segments = []
        split_parts = re.split(r'(")', text_parts)  # Split by quote marks, keeping the quotes

        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)
                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
                if audio_path:
                    audio_segments.append(audio_path)

        return start_time_ms, audio_segments
    return None, None

async def transcript_to_speech(transcript_text, voice, rate, pitch):
    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
    previous_end_time_ms = 0
    previous_start_time_ms = 0

    for i, line in enumerate(lines):
        start_time, audio_paths = await process_transcript_line(line, voice, rate, pitch)
        if start_time is not None and audio_paths:
            combined_line_audio = AudioSegment.empty()
            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 combined_line_audio:
                current_audio_duration = len(combined_line_audio)
                intended_start_time = start_time

                if i > 0:
                    time_difference = start_time - previous_start_time_ms
                    if current_audio_duration > time_difference:
                        intended_start_time = previous_end_time_ms

                timed_audio_segments.append({'start': intended_start_time, 'audio': combined_line_audio})
                previous_start_time_ms = start_time
                previous_end_time_ms = max(previous_end_time_ms, intended_start_time + current_audio_duration)
                max_end_time_ms = max(max_end_time_ms, previous_end_time_ms)
        elif audio_paths:
            for path in audio_paths:
                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))
    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 "1F Different Voice"
    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!"
    ```
    ***************************************************************************************************
    1M = en-AU-WilliamNeural - en-AU (Male)
    1F = en-GB-SoniaNeural - en-GB (Female)
    2M = en-GB-RyanNeural - en-GB (Male)
    2F = en-US-JennyNeural - en-US (Female)
    3M = en-US-BrianMultilingualNeural - en-US (Male)
    3F = en-HK-YanNeural - en-HK (Female)
    4M = en-GB-ThomasNeural - en-GB (Male)
    4F = en-US-EmmaNeural - en-US (Female)
    1O = en-GB-RyanNeural - en-GB (Male) # Old Man
    1C = en-GB-MaisieNeural - en-GB (Female) # Child
    1V = vi-VN-HoaiMyNeural - vi-VN (Female) # Vietnamese (Female)
    2V = vi-VN-NamMinhNeural - vi-VN (Male) # Vietnamese (Male)
    3V = vi-VN-HoaiMyNeural - vi-VN (Female) # Vietnamese (Female)
    4V = vi-VN-NamMinhNeural - vi-VN (Male) # Vietnamese (Male)
    ****************************************************************************************************
    """
    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.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) # Removed the duplicate value argument
        ],
        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",
        description=description,
        analytics_enabled=False,
        allow_flagging=False
    )
    return demo

if __name__ == "__main__":
    demo = asyncio.run(create_demo())
    demo.launch()