import spaces import gradio as gr import edge_tts import asyncio import tempfile import os import re # Import the regular expression module 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} # Text-to-speech function for a single paragraph with SS handling async def paragraph_to_speech(text, voice, rate, pitch): voice3 ="en-US-BrianMultilingualNeural - en-US (Male)" #good for reading voice1F ="en-US-EmmaNeural - en-US (Female)" voice2 = "it-IT-GiuseppeMultilingualNeural - it-IT (Male)" voice2F = "en-US-JennyNeural - en-US (Female)" voice1 = "en-AU-WilliamNeural - en-AU (Male)" voice3F = "en-HK-YanNeural - en-HK (Female)" voice4 = "en-GB-MaisieNeural - en-GB (Female)" #Child voice5 = "en-GB-RyanNeural - en-GB (Male)" #Old Man if not text.strip(): return None, [] # Return None for audio path and empty list for silence audio_segments = [] silence_durations = [] parts = re.split(r'(SS\d+\.?\d*)', text) for part in parts: if re.match(r'SS\d+\.?\d*', part): # At the top of your file: #SILENCE_PATH = Path(__file__).parent.absolute() / "Silence.mp3" # At the top of your file (assuming you uploaded "Silence.mp3" to root) #SILENCE_PATH = Path(__file__).parent.absolute() / "Silence.mp3" # At the top of your file: #SILENCE_PATH = Path(__file__).parent.absolute() / "static" / "intro.mp3" #if SILENCE_PATH.exists(): # audio_segments.append(str(SILENCE_PATH)) # print(f"Silence.mp3 file found at {SILENCE_PATH} and is inserted") #else: silence_duration = float(part[2:]) * 1000 # Convert to milliseconds print(f"Silence.mp3 file NOT FOUND") silence_file_path = get_silence(silence_duration) # Store the returned filename audio_segments.append(silence_file_path) # Use the stored filename elif part.strip(): processed_text = part current_voice = voice current_rate = rate current_pitch = pitch if part.startswith("1F"): processed_text = part[2:] current_voice = voice1F.split(" - ")[0] elif part.startswith("2F"): processed_text = part[2:] current_voice = voice2F.split(" - ")[0] elif part.startswith("3F"): processed_text = part[2:] current_voice = voice3F.split(" - ")[0] elif part.startswith("1M"): processed_text = part[2:] current_voice = voice1.split(" - ")[0] elif part.startswith("2M"): processed_text = part[2:] current_voice = voice2.split(" - ")[0] elif part.startswith("3M"): processed_text = part[2:] current_voice = voice3.split(" - ")[0] elif part.startswith("1C"): processed_text = part[2:] current_voice = voice4.split(" - ")[0] elif part.startswith("1O"): processed_text = part[2:] current_voice = voice5.split(" - ")[0] current_pitch = -30 current_rate = -20 else: # Use selected voice, or fallback to default #voice_short_name = (voice or default_voice).split(" - ")[0] current_voice = (voice or default_voice).split(" - ")[0] processed_text=part[:] rate_str = f"{current_rate:+d}%" pitch_str = f"{current_pitch:+d}Hz" communicate = edge_tts.Communicate(processed_text, current_voice, rate=rate_str, pitch=pitch_str) with tempfile.NamedTemporaryFile(delete=False, suffix=".mp3") as tmp_file: tmp_path = tmp_file.name await communicate.save(tmp_path) audio_segments.append(tmp_path) else: audio_segments.append(None) # Empty string return audio_segments, silence_durations # Main text-to-speech function that processes paragraphs and silence async def text_to_speech(text, voice, rate, pitch): if not text.strip(): return None, gr.Warning("Please enter text to convert.") if not voice: return None, gr.Warning("Please select a voice.") paragraphs = [p.strip() for p in re.split(r'"', text) if p.strip()] final_audio_segments = [] for paragraph in paragraphs: audio_paths, silence_times = await paragraph_to_speech(paragraph, voice, rate, pitch) if audio_paths: for i, path in enumerate(audio_paths): final_audio_segments.append(path) if i < len(silence_times): final_audio_segments.append(silence_times[i]) if not any(isinstance(item, str) for item in final_audio_segments): return None, None # No actual audio generated if all(not isinstance(item, str) for item in final_audio_segments): return None, "Only silence markers found." combined_audio_path = tempfile.mktemp(suffix=".mp3") with open(combined_audio_path, 'wb') as outfile: for segment in final_audio_segments: if isinstance(segment, str): try: with open(segment, 'rb') as infile: outfile.write(infile.read()) os.remove(segment) # Clean up individual files except FileNotFoundError: print(f"Warning: Audio file not found: {segment}") return combined_audio_path, None # Gradio interface function @spaces.GPU def tts_interface(text, voice, rate, pitch): audio, warning = asyncio.run(text_to_speech(text, voice, rate, pitch)) return audio, warning # Create Gradio application import gradio as gr async def create_demo(): voices = await get_voices() default_voice = "en-US-AndrewMultilingualNeural - en-US (Male)" # 👈 Pick one of the available voices description = """ Default = male, other voices 1F:US_Emma, 2F:US_Jenny, 3F:HK_Yan, 1M:AU_Will, 2M:IT_Guiseppe,3M:US_Brian, 1C: Childvoice, 1O = OldMan You can insert silence using the marker 'SS' (This will insert a Silence period from the Silence.mp3 file). Enter your text, select a voice, and adjust the speech rate and pitch. The application will process your text paragraph by paragraph (separated by two blank lines). """ demo = gr.Interface( fn=tts_interface, inputs=[ gr.Textbox(label="Input Text", lines=5, placeholder="Separate paragraphs with two blank lines. Use 'SS[duration]' for silence."), gr.Dropdown(choices=[""] + list(voices.keys()), label="Select 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) ], outputs=[ gr.Audio(label="Generated Audio", type="filepath"), gr.Markdown(label="Warning", visible=False) ], title="Voicecloning.be Text-to-Speech with Silence Insertion (Paragraph by Paragraph)", description=description, article="Process text paragraph by paragraph for smoother output and insert silence markers.", analytics_enabled=False, allow_flagging=False ) return demo # Run the application if __name__ == "__main__": demo = asyncio.run(create_demo()) demo.launch()