import gradio as gr import torch import torchaudio import re import os from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan from speechbrain.pretrained import EncoderClassifier import numpy as np # --- Configuration --- device = "cuda" if torch.cuda.is_available() else "cpu" # --- ADD ALL YOUR VOICE FILES HERE --- # The code will automatically create a dropdown for these files. # Make sure these files are in the same directory as your script. VOICE_SAMPLE_FILES = ["46.wav", "90.wav", "150.wav", "355.wav"] # Directory to store speaker embedding files EMBEDDING_DIR = "speaker_embeddings" os.makedirs(EMBEDDING_DIR, exist_ok=True) # --- Load Models --- # This part loads all the necessary AI models. try: print("Loading models... This may take a moment.") processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts") model = SpeechT5ForTextToSpeech.from_pretrained("Somalitts/5aad").to(device) vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device) speaker_model = EncoderClassifier.from_hparams( source="speechbrain/spkrec-xvect-voxceleb", run_opts={"device": device}, savedir=os.path.join("pretrained_models", "spkrec-xvect-voxceleb") ) print("Models loaded successfully.") except Exception as e: raise gr.Error(f"Error loading models: {e}. Check your internet connection.") # A dictionary to cache loaded speaker embeddings in memory speaker_embeddings_cache = {} # --- Function to Get or Create Speaker Embedding --- def get_speaker_embedding(wav_file_path): """ Loads a speaker embedding from cache or file. If not found, creates and saves it. """ # Check cache first if wav_file_path in speaker_embeddings_cache: return speaker_embeddings_cache[wav_file_path] embedding_path = os.path.join(EMBEDDING_DIR, f"{os.path.basename(wav_file_path)}.pt") if os.path.exists(embedding_path): print(f"Loading existing embedding for {wav_file_path}") embedding = torch.load(embedding_path, map_location=device) speaker_embeddings_cache[wav_file_path] = embedding return embedding print(f"Creating new speaker embedding for {wav_file_path}...") if not os.path.exists(wav_file_path): raise gr.Error(f"Audio file not found: {wav_file_path}. Please make sure it's in the correct directory.") try: audio, sr = torchaudio.load(wav_file_path) if sr != 16000: audio = torchaudio.functional.resample(audio, sr, 16000) if audio.shape[0] > 1: audio = torch.mean(audio, dim=0, keepdim=True) with torch.no_grad(): embedding = speaker_model.encode_batch(audio.to(device)) embedding = torch.nn.functional.normalize(embedding, dim=2).squeeze() torch.save(embedding.cpu(), embedding_path) speaker_embeddings_cache[wav_file_path] = embedding.to(device) print(f"Embedding created and saved for {wav_file_path}.") return embedding.to(device) except Exception as e: raise gr.Error(f"Could not process audio file {wav_file_path}. Is it a valid WAV file? Error: {e}") # --- Text Processing Functions (Somali Number Conversion) --- # These functions remain the same. number_words = { 0: "eber", 1: "kow", 2: "labo", 3: "saddex", 4: "afar", 5: "shan", 6: "lix", 7: "toddobo", 8: "siddeed", 9: "sagaal", 10: "toban", 11: "kow iyo toban", 12: "labo iyo toban", 13: "saddex iyo toban", 14: "afar iyo toban", 15: "shan iyo toban", 16: "lix iyo toban", 17: "toddobo iyo toban", 18: "siddeed iyo toban", 19: "sagaal iyo toban", 20: "labaatan", 30: "soddon", 40: "afartan", 50: "konton", 60: "lixdan", 70: "toddobaatan", 80: "siddeetan", 90: "sagaashan", 100: "boqol", 1000: "kun", } def number_to_words(n): if n in number_words: return number_words[n] if n < 100: return number_words[n//10 * 10] + (" iyo " + number_words[n%10] if n%10 else "") if n < 1000: return (number_words[n//100] + " boqol" if n//100 > 1 else "boqol") + (" iyo " + number_to_words(n%100) if n%100 else "") if n < 1000000: return (number_to_words(n//1000) + " kun" if n//1000 > 1 else "kun") + (" iyo " + number_to_words(n%1000) if n%1000 else "") return str(n) def replace_numbers_with_words(text): return re.sub(r'\b\d+\b', lambda m: number_to_words(int(m.group())), text) def normalize_text(text): text = text.lower() text = replace_numbers_with_words(text) text = re.sub(r'[^\w\s\']', '', text) return text # --- Main Text-to-Speech Function --- def text_to_speech(text, voice_choice): """ Takes text and the chosen voice file, and returns audio. """ if not text: gr.Warning("Please enter some text.") return None, None if not voice_choice: gr.Warning("Please select a voice from the dropdown.") return None, None # Get the correct speaker embedding for the chosen voice speaker_embedding = get_speaker_embedding(voice_choice) normalized_text = normalize_text(text) inputs = processor(text=normalized_text, return_tensors="pt").to(device) with torch.no_grad(): speech = model.generate_speech( inputs["input_ids"], speaker_embeddings=speaker_embedding.unsqueeze(0), vocoder=vocoder ) return (16000, speech.cpu().numpy()) # --- Gradio Interface --- # The user interface now includes a dropdown menu for voice selection. iface = gr.Interface( fn=text_to_speech, inputs=[ gr.Textbox(label="Geli qoraalka af-Soomaaliga (Enter Somali Text)"), gr.Dropdown( VOICE_SAMPLE_FILES, label="Select Voice", info="Choose the voice you want to use for the speech.", value=VOICE_SAMPLE_FILES[0] if VOICE_SAMPLE_FILES else None # Default to the first voice ) ], outputs=gr.Audio(label="Codka La Abuuray (Generated Voice)", type="numpy"), title="Multi-Voice Somali Text-to-Speech", description="Enter Somali text, choose a voice from the dropdown, and click submit to generate speech.", examples=[ ["Sidee tahay saaxiib? Maanta waa maalin wanaagsan.", VOICE_SAMPLE_FILES[0]], ["Nabad gelyo, is arag dambe.", VOICE_SAMPLE_FILES[1] if len(VOICE_SAMPLE_FILES) > 1 else VOICE_SAMPLE_FILES[0]], ] ) # Launch the web interface if __name__ == "__main__": # Pre-load embeddings for a faster startup experience print("Pre-loading all voice embeddings...") for voice_file in VOICE_SAMPLE_FILES: get_speaker_embedding(voice_file) print("All voices are ready. Launching interface.") iface.launch(share=True)