Update app.py

app.py

@@ -19,19 +19,16 @@ processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
 model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts")
 vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
 
-# Max tokens allowed by model
-MAX_TOKENS = 600
-
 # Map integer to speaker embedding URL
 SPEAKER_EMBEDDINGS = {
-    0: "https://huggingface.co/datasets/Xenova/transformers.js-docs/resolve/main/speaker_embeddings.bin",
-    1: "https://huggingface.co/datasets/Xenova/cmu-arctic-xvectors-extracted/resolve/main/cmu_us_slt_arctic-wav-arctic_a0001.bin",
-    2: "https://huggingface.co/datasets/Xenova/cmu-arctic-xvectors-extracted/resolve/main/cmu_us_clb_arctic-wav-arctic_a0001.bin",
-    3: "https://huggingface.co/datasets/Xenova/cmu-arctic-xvectors-extracted/resolve/main/cmu_us_bdl_arctic-wav-arctic_a0003.bin",
-    4: "https://huggingface.co/datasets/Xenova/cmu-arctic-xvectors-extracted/resolve/main/cmu_us_rms_arctic-wav-arctic_a0003.bin",
-    5: "https://huggingface.co/datasets/Xenova/cmu-arctic-xvectors-extracted/resolve/main/cmu_us_jmk_arctic-wav-arctic_a0002.bin",
-    6: "https://huggingface.co/datasets/Xenova/cmu-arctic-xvectors-extracted/resolve/main/cmu_us_awb_arctic-wav-arctic_b0002.bin",
-    7: "https://huggingface.co/datasets/Xenova/cmu-arctic-xvectors-extracted/resolve/main/cmu_us_ksp_arctic-wav-arctic_a0007.bin",
+    0: "https://huggingface.co/datasets/Xenova/transformers.js-docs/resolve/main/speaker_embeddings.bin",  # Normal
+    1: "https://huggingface.co/datasets/Xenova/cmu-arctic-xvectors-extracted/resolve/main/cmu_us_slt_arctic-wav-arctic_a0001.bin",  # US female 1
+    2: "https://huggingface.co/datasets/Xenova/cmu-arctic-xvectors-extracted/resolve/main/cmu_us_clb_arctic-wav-arctic_a0001.bin",  # US female 2
+    3: "https://huggingface.co/datasets/Xenova/cmu-arctic-xvectors-extracted/resolve/main/cmu_us_bdl_arctic-wav-arctic_a0003.bin",  # US male 1
+    4: "https://huggingface.co/datasets/Xenova/cmu-arctic-xvectors-extracted/resolve/main/cmu_us_rms_arctic-wav-arctic_a0003.bin",  # US male 2
+    5: "https://huggingface.co/datasets/Xenova/cmu-arctic-xvectors-extracted/resolve/main/cmu_us_jmk_arctic-wav-arctic_a0002.bin",  # Canadian male
+    6: "https://huggingface.co/datasets/Xenova/cmu-arctic-xvectors-extracted/resolve/main/cmu_us_awb_arctic-wav-arctic_b0002.bin",  # Scottish male
+    7: "https://huggingface.co/datasets/Xenova/cmu-arctic-xvectors-extracted/resolve/main/cmu_us_ksp_arctic-wav-arctic_a0007.bin",  # Indian male
 }
 
 
@@ -42,7 +39,7 @@ def load_speaker_embedding(url: str) -> torch.Tensor:
     return embedding.unsqueeze(0)
 
 
-def smooth_audio(audio: np.ndarray, window_size: int) -> np.ndarray:
+def smooth_audio_nodejs(audio: np.ndarray, window_size: int) -> np.ndarray:
     smoothed = np.copy(audio)
     half = window_size // 2
     for i in range(len(audio)):
@@ -57,26 +54,6 @@ def apply_fade_out(audio: np.ndarray, fade_samples: int = 256) -> np.ndarray:
     audio[-fade_samples:] *= fade
     return audio
 
-
-def chunk_text(text: str, processor, max_tokens: int = MAX_TOKENS):
-    words = text.split()
-    chunks = []
-    current_chunk = []
-
-    for word in words:
-        current_chunk.append(word)
-        tokens = processor(" ".join(current_chunk), return_tensors="pt")["input_ids"]
-        if tokens.size(1) > max_tokens:
-            current_chunk.pop()
-            chunks.append(" ".join(current_chunk))
-            current_chunk = [word]
-
-    if current_chunk:
-        chunks.append(" ".join(current_chunk))
-
-    return chunks
-
-
 @app.get("/speak")
 def speak(
     text: str = Query(..., description="Text to convert to speech"),
@@ -85,17 +62,9 @@ def speak(
     embedding_url = SPEAKER_EMBEDDINGS[speaker]
     speaker_embedding = load_speaker_embedding(embedding_url)
 
-    chunks = chunk_text(text, processor, MAX_TOKENS)
-    audio_list = []
-
-    for chunk in chunks:
-        inputs = processor(text=chunk, return_tensors="pt")
-        speech = model.generate_speech(inputs["input_ids"], speaker_embedding, vocoder=vocoder)
-        audio_chunk = speech.numpy().astype(np.float32)
-        audio_list.append(audio_chunk)
-
-    # Concatenate all chunks
-    audio = np.concatenate(audio_list)
+    inputs = processor(text=text, return_tensors="pt")
+    speech = model.generate_speech(inputs["input_ids"], speaker_embedding, vocoder=vocoder)
+    audio = speech.numpy().astype(np.float32)
 
     # --- Normalize ---
     peak = np.max(np.abs(audio))
@@ -103,10 +72,10 @@ def speak(
         audio = (audio / peak) * NORMALIZATION_LEVEL
 
     # --- Smooth audio ---
-    audio = smooth_audio(audio, SMOOTHING_WINDOW * 2)
+    audio = smooth_audio_nodejs(audio, SMOOTHING_WINDOW * 2)  # Slightly larger window
 
-    # --- Fade out ---
-    fade_samples = min(512, len(audio)//10)
+    # --- Fade out to remove clicks at the end ---
+    fade_samples = min(512, len(audio)//10)  # ~30ms at 16kHz
     audio = apply_fade_out(audio, fade_samples)
 
     # --- Bit depth ---