Update audiobook.py

audiobook.py

@@ -1,188 +1,195 @@
-import os
-import re
-import torch
-import numpy as np
-from scipy.io.wavfile import write
-from tts import commons
-from tts import utils
-from tts.models import SynthesizerTrn
-from text.symbols import symbols
-from text import text_to_sequence
-from phonemizer.backend.espeak.wrapper import EspeakWrapper
-from safetensors.torch import load_file
-
-_ESPEAK_LIBRARY = r"C:\Program Files\eSpeak NG\libespeak-ng.dll"
-if os.path.exists(_ESPEAK_LIBRARY):
-    EspeakWrapper.set_library(_ESPEAK_LIBRARY)
-    print(f"✅ Found eSpeak-ng: {_ESPEAK_LIBRARY}")
-else:
-    print("⚠️ eSpeak-ng not found (ok if already working)")
-
-
-MODEL_PATH = "checkpoints/sonya-tts.safetensors"
-CONFIG_PATH = "checkpoints/config.json"
-
-OUTPUT_WAV_SHORT = "output.wav"
-OUTPUT_WAV_LONG  = "audiobook.wav"
-
-USE_LONG_FORM = True  # ← change to False for short text
-
-TEXT = """
-You’re 100 percent right. LeetCode solutions should be simple, readable,
-and teachable. This long form inference mode allows narration of long
-paragraphs without instability. The audio remains smooth and natural.
-without breaking audio quality. It works smoothly. The Problem: VITS crashes or creates garbage audio if you feed it a whole paragraph or page of text. The Solution: A script that automatically splits text into sentences, generates audio for each, and stitches them together with natural pauses.
-Why it stands out: It allows you to convert entire PDF chapters or articles into audio files automatically. Since the architectural changes for zero-shot cloning adding embeddings require retraining, let’s focus on inference-side features you can add right now to your existing G_10400.pth model.
-These features will turn your simple "Text-to-WAV" script into a powerful Audiobook & Content Creation Tool.
-"""
-
-def save_wav_int16(path, audio, sample_rate):
-    audio = np.clip(audio, -1.0, 1.0)
-    audio = (audio * 32767).astype(np.int16)
-    write(path, sample_rate, audio)
-
-
-def clean_text_for_vits(text):
-    text = text.strip()
-
-   
-    text = text.replace("’", "'")
-    text = text.replace("“", '"').replace("”", '"')
-    text = text.replace("–", "-").replace("—", "-")
-
-  
-    text = re.sub(r"[()\[\]{}<>]", "", text)
-
-   
-    text = re.sub(r"[^a-zA-Z0-9\s.,!?'\-]", "", text)
-
-    text = re.sub(r"\s+", " ", text)
-
-    return text
-
-def get_text(text, hps):
-    text = clean_text_for_vits(text)
-    text_norm = text_to_sequence(text, hps.data.text_cleaners)
-    if hps.data.add_blank:
-        text_norm = commons.intersperse(text_norm, 0)
-    return torch.LongTensor(text_norm)
-
-
-def split_sentences(text):
-    text = clean_text_for_vits(text)
-    if not text:
-        return []
-    return re.split(r'(?<=[.!?])\s+', text)
-
-
-def generate_audiobook(
-    net_g,
-    hps,
-    text,
-    device,
-    output_file,
-    noise_scale=0.5,
-    noise_scale_w=0.6,
-    length_scale=1.0,
-    base_pause=0.4,
-):
-    print("📖 Long-form audiobook mode enabled")
-
-    sentences = split_sentences(text)
-    print(f"🔹 Sentences: {len(sentences)}")
-
-    audio_chunks = []
-
-    for i, sent in enumerate(sentences):
-        sent = sent.strip()
-        if not sent:
-            continue
-
-        stn_tst = get_text(sent, hps)
-
-        with torch.no_grad():
-            x = stn_tst.to(device).unsqueeze(0)
-            x_len = torch.LongTensor([stn_tst.size(0)]).to(device)
-
-            audio = net_g.infer(
-                x,
-                x_len,
-                noise_scale=noise_scale,
-                noise_scale_w=noise_scale_w,
-                length_scale=length_scale,
-            )[0][0, 0].cpu().numpy()
-
-        if sent.endswith("?"):
-            pause = base_pause + 0.15
-        elif sent.endswith("!"):
-            pause = base_pause
-        else:
-            pause = base_pause + 0.05
-
-        silence = np.zeros(int(hps.data.sampling_rate * pause))
-
-        audio_chunks.append(audio)
-        audio_chunks.append(silence)
-
-        print(f"   ✅ Sentence {i+1}/{len(sentences)} done")
-
-    final_audio = np.concatenate(audio_chunks)
-    save_wav_int16(output_file, final_audio, hps.data.sampling_rate)
-
-    print(f"🎉 Audiobook saved: {os.path.abspath(output_file)}")
-
-
-def main():
-    if not os.path.exists(CONFIG_PATH):
-        print("❌ Config file not found")
-        return
-
-    hps = utils.get_hparams_from_file(CONFIG_PATH)
-
-    device = "cuda" if torch.cuda.is_available() else "cpu"
-    print(f"🚀 Using device: {device}")
-
-    # Load model
-    net_g = SynthesizerTrn(
-        len(symbols),
-        hps.data.filter_length // 2 + 1,
-        hps.train.segment_size // hps.data.hop_length,
-        **hps.model,
-    ).to(device)
-    net_g.eval()
-
-    # Load checkpoint
-    state_dict = load_file(MODEL_PATH, device=device)
-    net_g.load_state_dict(state_dict)
-    print(f"✅ Loaded model: {MODEL_PATH}")
-
-    
-    if USE_LONG_FORM:
-        generate_audiobook(
-            net_g,
-            hps,
-            TEXT,
-            device,
-            OUTPUT_WAV_LONG,
-        )
-    else:
-        print("🗣️ Short-text inference")
-
-        stn_tst = get_text(TEXT, hps)
-        with torch.no_grad():
-            x = stn_tst.to(device).unsqueeze(0)
-            x_len = torch.LongTensor([stn_tst.size(0)]).to(device)
-
-            audio = net_g.infer(
-                x,
-                x_len,
-                noise_scale=0.5,
-                noise_scale_w=0.6,
-                length_scale=1.0,
-            )[0][0, 0].cpu().numpy()
-
-        save_wav_int16(OUTPUT_WAV_SHORT, audio, hps.data.sampling_rate)
-        print(f"💾 Saved audio: {os.path.abspath(OUTPUT_WAV_SHORT)}")
-
-if __name__ == "__main__":
-    main()
+import os
+import re
+import torch
+import numpy as np
+from scipy.io.wavfile import write
+from tts import commons
+from tts import utils
+from tts.models import SynthesizerTrn
+from text.symbols import symbols
+from text import text_to_sequence
+from phonemizer.backend.espeak.wrapper import EspeakWrapper
+from safetensors.torch import load_file
+from huggingface_hub import hf_hub_download
+
+_ESPEAK_LIBRARY = r"C:\Program Files\eSpeak NG\libespeak-ng.dll"
+if os.path.exists(_ESPEAK_LIBRARY):
+    EspeakWrapper.set_library(_ESPEAK_LIBRARY)
+    print(f"✅ Found eSpeak-ng: {_ESPEAK_LIBRARY}")
+else:
+    print("⚠️ eSpeak-ng not found (ok if already working)")
+
+
+REPO_ID = "PatnaikAshish/Sonya-TTS"
+MODEL_FILENAME = "sonya-tts.safetensors"
+CONFIG_FILENAME = "config.json"
+
+LOCAL_MODEL_PATH = "checkpoints/sonya-tts.safetensors"
+LOCAL_CONFIG_PATH = "checkpoints/config.json"
+OUTPUT_WAV_SHORT = "output.wav"
+OUTPUT_WAV_LONG  = "audiobook.wav"
+
+USE_LONG_FORM = True  # ← change to False for short text
+
+TEXT = """
+A neural network or Artificial Neural Network is a computer system inspired by the human brain, 
+using interconnected nodes neurons in layers to recognize complex patterns in data for tasks like 
+image recognition, language processing, and prediction
+"""
+
+def save_wav_int16(path, audio, sample_rate):
+    audio = np.clip(audio, -1.0, 1.0)
+    audio = (audio * 32767).astype(np.int16)
+    write(path, sample_rate, audio)
+
+
+def clean_text_for_vits(text):
+    text = text.strip()
+
+   
+    text = text.replace("’", "'")
+    text = text.replace("“", '"').replace("”", '"')
+    text = text.replace("–", "-").replace("—", "-")
+
+  
+    text = re.sub(r"[()\[\]{}<>]", "", text)
+
+   
+    text = re.sub(r"[^a-zA-Z0-9\s.,!?'\-]", "", text)
+
+    text = re.sub(r"\s+", " ", text)
+
+    return text
+
+def get_text(text, hps):
+    text = clean_text_for_vits(text)
+    text_norm = text_to_sequence(text, hps.data.text_cleaners)
+    if hps.data.add_blank:
+        text_norm = commons.intersperse(text_norm, 0)
+    return torch.LongTensor(text_norm)
+
+
+def split_sentences(text):
+    text = clean_text_for_vits(text)
+    if not text:
+        return []
+    return re.split(r'(?<=[.!?])\s+', text)
+
+
+def generate_audiobook(
+    net_g,
+    hps,
+    text,
+    device,
+    output_file,
+    noise_scale=0.5,
+    noise_scale_w=0.6,
+    length_scale=1.0,
+    base_pause=0.4,
+):
+    print("📖 Long-form audiobook mode enabled")
+
+    sentences = split_sentences(text)
+    print(f"🔹 Sentences: {len(sentences)}")
+
+    audio_chunks = []
+
+    for i, sent in enumerate(sentences):
+        sent = sent.strip()
+        if not sent:
+            continue
+
+        stn_tst = get_text(sent, hps)
+
+        with torch.no_grad():
+            x = stn_tst.to(device).unsqueeze(0)
+            x_len = torch.LongTensor([stn_tst.size(0)]).to(device)
+
+            audio = net_g.infer(
+                x,
+                x_len,
+                noise_scale=noise_scale,
+                noise_scale_w=noise_scale_w,
+                length_scale=length_scale,
+            )[0][0, 0].cpu().numpy()
+
+        if sent.endswith("?"):
+            pause = base_pause + 0.15
+        elif sent.endswith("!"):
+            pause = base_pause
+        else:
+            pause = base_pause + 0.05
+
+        silence = np.zeros(int(hps.data.sampling_rate * pause))
+
+        audio_chunks.append(audio)
+        audio_chunks.append(silence)
+
+        print(f"   ✅ Sentence {i+1}/{len(sentences)} done")
+
+    final_audio = np.concatenate(audio_chunks)
+    save_wav_int16(output_file, final_audio, hps.data.sampling_rate)
+
+    print(f"🎉 Audiobook saved: {os.path.abspath(output_file)}")
+
+
+def main():
+    if os.path.exists(LOCAL_MODEL_PATH) and os.path.exists(LOCAL_CONFIG_PATH):
+        print("✅ Loading Sonya TTS from local checkpoints...")
+        model_path = LOCAL_MODEL_PATH
+        config_path = LOCAL_CONFIG_PATH
+    else:
+        print("🌍 Downloading Sonya TTS from Hugging Face...")
+        model_path = hf_hub_download(repo_id=REPO_ID, filename=MODEL_FILENAME)
+        config_path = hf_hub_download(repo_id=REPO_ID, filename=CONFIG_FILENAME)
+
+
+    hps = utils.get_hparams_from_file(config_path)
+
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    print(f"🚀 Using device: {device}")
+
+    # Load model
+    net_g = SynthesizerTrn(
+        len(symbols),
+        hps.data.filter_length // 2 + 1,
+        hps.train.segment_size // hps.data.hop_length,
+        **hps.model,
+    ).to(device)
+    net_g.eval()
+
+    # Load checkpoint
+    state_dict = load_file(model_path)
+    net_g.load_state_dict(state_dict)
+    print(f"✅ Loaded model: {model_path}")
+
+    
+    if USE_LONG_FORM:
+        generate_audiobook(
+            net_g,
+            hps,
+            TEXT,
+            device,
+            OUTPUT_WAV_LONG,
+        )
+    else:
+        print("🗣️ Short-text inference")
+
+        stn_tst = get_text(TEXT, hps)
+        with torch.no_grad():
+            x = stn_tst.to(device).unsqueeze(0)
+            x_len = torch.LongTensor([stn_tst.size(0)]).to(device)
+
+            audio = net_g.infer(
+                x,
+                x_len,
+                noise_scale=0.5,
+                noise_scale_w=0.6,
+                length_scale=1.0,
+            )[0][0, 0].cpu().numpy()
+
+        save_wav_int16(OUTPUT_WAV_SHORT, audio, hps.data.sampling_rate)
+        print(f"💾 Saved audio: {os.path.abspath(OUTPUT_WAV_SHORT)}")
+
+if __name__ == "__main__":
+    main()