Improve TTS: sentence-aware split, error handling, text truncation, cleanup

app.py

@@ -1,63 +1,100 @@
 import gradio as gr
 from kokoro import KPipeline
-import soundfile as sf
-import torch
 import numpy as np
 
-# Inicialização das Pipelines por idioma
-# 'a' = American English, 'b' = British English, 'p' = Brazilian Portuguese
+# Inicialização das Pipelines por idioma (carregadas uma vez ao iniciar)
+# a = American English, b = British English, p = Brazilian Portuguese
 pipelines = {
-    'a': KPipeline(lang_code='a'),
-    'b': KPipeline(lang_code='b'),
-    'p': KPipeline(lang_code='p')
+    "a": KPipeline(lang_code="a"),
+    "b": KPipeline(lang_code="b"),
+    "p": KPipeline(lang_code="p"),
 }
 
-def get_pipeline_for_voice(voice_id):
-    if voice_id.startswith(('af_', 'am_')):
-        return pipelines['a']
-    elif voice_id.startswith(('bf_', 'bm_')):
-        return pipelines['b']
-    elif voice_id.startswith(('pf_', 'pm_')):
-        return pipelines['p']
-    return pipelines['a']
-
-def generate_tts(text, voice_id, speed):
-    pipeline = get_pipeline_for_voice(voice_id)
-    generator = pipeline(
-        text, voice=voice_id,
-        speed=speed, split_pattern=r'\n+'
-    )
-    
-    full_audio = []
-    for i, (gs, ps, audio) in enumerate(generator):
-        full_audio.append(audio)
-    
-    if not full_audio:
+
+def get_pipeline_for_voice(voice_id: str) -> KPipeline:
+    if voice_id.startswith(("af_", "am_")):
+        return pipelines["a"]
+    elif voice_id.startswith(("bf_", "bm_")):
+        return pipelines["b"]
+    elif voice_id.startswith(("pf_", "pm_")):
+        return pipelines["p"]
+    return pipelines["a"]
+
+
+def generate_tts(text: str, voice_id: str, speed: float):
+    """
+    Gera áudio TTS a partir de texto usando Kokoro-82M.
+    Retorna tupla (sample_rate, audio_array) para gr.Audio.
+    """
+    try:
+        text = (text or "").strip()
+        if not text:
+            return None
+
+        # Trunca textos muito longos para evitar timeout
+        MAX_CHARS = 600
+        if len(text) > MAX_CHARS:
+            print(f"[TTS] Texto truncado: {len(text)} -> {MAX_CHARS} chars")
+            text = text[:MAX_CHARS]
+
+        # Textos curtos: sem split (evita pausas desnecessárias)
+        # Textos longos: split por fronteira de sentença para melhor prosódia
+        split_pattern = None if len(text) < 80 else r"(?<=[.!?…])\s+"
+
+        pipeline = get_pipeline_for_voice(voice_id)
+        generator = pipeline(
+            text,
+            voice=voice_id,
+            speed=speed,
+            split_pattern=split_pattern,
+        )
+
+        segments = []
+        for _gs, _ps, audio in generator:
+            seg = np.array(audio, dtype=np.float32)
+            if seg.ndim > 1:          # garante mono
+                seg = seg.mean(axis=-1)
+            if seg.size > 0:
+                segments.append(seg)
+
+        if not segments:
+            print("[TTS] Nenhum segmento gerado.")
+            return None
+
+        combined = np.concatenate(segments).astype(np.float32)
+        return (24000, combined)
+
+    except Exception as e:
+        print(f"[TTS] Erro: {e}")
         return None
-        
-    combined_audio = np.concatenate(full_audio)
-    return (24000, combined_audio)
 
-# Lista de vozes conforme documentação kokoro-82M
+
+# Vozes disponíveis (Kokoro-82M)
 VOICES = [
-    "af_heart", "af_alloy", "af_aoede", "af_bella", "af_jessica", 
+    # American English — female
+    "af_heart", "af_alloy", "af_aoede", "af_bella", "af_jessica",
     "af_kore", "af_nicole", "af_nova", "af_river", "af_sarah", "af_sky",
-    "am_adam", "am_echo", "am_eric", "am_fenrir", "am_liam", "am_michael", 
-    "am_onyx", "am_orion", "am_puck",
-    "bf_alice", "bf_emma", "bf_isabella", "bm_george", "bm_lewis",
-    "pf_dora", "pm_alex", "pm_santa"
+    # American English — male
+    "am_adam", "am_echo", "am_eric", "am_fenrir", "am_liam",
+    "am_michael", "am_onyx", "am_orion", "am_puck",
+    # British English — female / male
+    "bf_alice", "bf_emma", "bf_isabella",
+    "bm_george", "bm_lewis",
+    # Brazilian Portuguese — female / male
+    "pf_dora", "pm_alex", "pm_santa",
 ]
 
 demo = gr.Interface(
     fn=generate_tts,
     inputs=[
-        gr.Textbox(label="Texto", lines=3),
+        gr.Textbox(label="Texto", lines=3, placeholder="Digite o texto aqui..."),
         gr.Dropdown(choices=VOICES, value="af_heart", label="Voz"),
-        gr.Slider(minimum=0.5, maximum=2.0, value=1.0, step=0.1, label="Velocidade")
+        gr.Slider(minimum=0.5, maximum=2.0, value=1.0, step=0.1, label="Velocidade"),
     ],
     outputs=gr.Audio(label="Áudio Gerado"),
-    title="Kokoro-82M Multi-Language TTS",
-    description="Suporte a Inglês (US/UK) e Português (BR)."
+    title="Kokoro-82M TTS",
+    description="Inglês Americano / Britânico e Português Brasileiro.",
+    cache_examples=False,
 )
 
 if __name__ == "__main__":