import gradio as gr import torch import torchaudio import re import os import numpy as np import scipy.io.wavfile from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan from speechbrain.pretrained import EncoderClassifier # --- Configuration --- device = "cuda" if torch.cuda.is_available() else "cpu" print(f"Using device: {device}") VOICE_SAMPLE_FILES = ["1.wav"] EMBEDDING_DIR = "speaker_embeddings" os.makedirs(EMBEDDING_DIR, exist_ok=True) # --- Load models --- try: print("Loading models...") processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts") model = SpeechT5ForTextToSpeech.from_pretrained("Somalitts/8aad").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}.") speaker_embeddings_cache = {} def get_speaker_embedding(wav_file_path): 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): embedding = torch.load(embedding_path, map_location=device) speaker_embeddings_cache[wav_file_path] = embedding return embedding if not os.path.exists(wav_file_path): raise gr.Error(f"Audio file not found: {wav_file_path}") 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) return embedding.to(device) except Exception as e: raise gr.Error(f"Could not process audio file {wav_file_path}. Error: {e}") # --- Number words dictionary and functions --- 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 < 1_000_000: return (number_to_words(n // 1000) + " kun" if n // 1000 > 1 else "kun") + ( " iyo " + number_to_words(n % 1000) if n % 1000 else "") if n < 1_000_000_000: return (number_to_words(n // 1_000_000) + " milyan" if n // 1_000_000 > 1 else "milyan") + ( " iyo " + number_to_words(n % 1_000_000) if n % 1_000_000 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 # --- Split long text into chunks by word count --- def split_long_text_into_chunks(text, max_words=18): words = text.split() chunks = [] for i in range(0, len(words), max_words): chunk = ' '.join(words[i:i + max_words]) chunks.append(chunk) return chunks # --- Main TTS function --- def text_to_speech(text, voice_choice): if not text or not voice_choice: gr.Warning("Fadlan geli qoraal oo dooro cod.") return None speaker_embedding = get_speaker_embedding(voice_choice) text_chunks = split_long_text_into_chunks(text) audio_chunks = [] for idx, chunk in enumerate(text_chunks): chunk = chunk.strip() if not chunk: continue norm_chunk = normalize_text(chunk) inputs = processor(text=norm_chunk, return_tensors="pt").to(device) with torch.no_grad(): speech = model.generate( input_ids=inputs["input_ids"], speaker_embeddings=speaker_embedding.unsqueeze(0), do_sample=True, top_k=50, temperature=0.75, repetition_penalty=1.2, max_new_tokens=512 ) audio = vocoder(speech).cpu().squeeze().numpy() audio_chunks.append(audio) # Pause after each chunk if idx < len(text_chunks) - 1: pause = np.zeros(int(16000 * 0.8)) # 0.8s pause audio_chunks.append(pause) final_audio = np.concatenate(audio_chunks) return (16000, final_audio) # --- Gradio Interface --- iface = gr.Interface( fn=text_to_speech, inputs=[ gr.Textbox(label="Geli qoraalka af-Soomaaliga (Enter Somali Text)", lines=7, placeholder="Qoraalka geli halkan..."), gr.Dropdown( VOICE_SAMPLE_FILES, label="Dooro Codka (Select Voice)", value=VOICE_SAMPLE_FILES[0] if VOICE_SAMPLE_FILES else None ) ], outputs=gr.Audio(label="Codka La Abuuray (Generated Audio)", type="numpy"), title="Multi-Voice Somali Text-to-Speech", description="Geli qoraal Soomaali ah, dooro cod, kadib riix 'Submit' si aad u abuurto hadal." ) # --- Launch App --- if __name__ == "__main__": if not all(os.path.exists(f) for f in VOICE_SAMPLE_FILES): raise FileNotFoundError("Fadlan hubi inaad faylasha codka ku dartay.") print("Diyaarinta codadka...") for voice_file in VOICE_SAMPLE_FILES: get_speaker_embedding(voice_file) print("Dhammaan waa diyaar. Barnaamijku wuu furmayaa.") iface.launch(share=True)