Update app.py

app.py

@@ -8,7 +8,6 @@ from pydub import AudioSegment
 from transformers import VitsModel, AutoTokenizer
 
 # ---------- Configuration --------------------------------------------------
-# Define available TTS models here. Add new entries as needed.
 TTS_MODELS = {
     "Ewe": {
         "tokenizer": "FarmerlineML/Ewe-tts-2025_v3",
@@ -30,13 +29,25 @@ TTS_MODELS = {
         "tokenizer": "FarmerlineML/luganda_TTS_v1",
         "checkpoint": "FarmerlineML/luganda_TTS_v1"
     },
-    
     "Yoruba": {
         "tokenizer": "FarmerlineML/yoruba_tts-2025",
         "checkpoint": "FarmerlineML/yoruba_tts-2025"
     },
 }
 
+# Per-language inference parameters
+# noise_scale:          prosody expressiveness (higher = more varied intonation)
+# noise_scale_duration: rhythm/timing variation (higher = more natural pausing)
+# speaking_rate:        speed (lower = slower, clearer)
+LANG_PARAMS = {
+    "Yoruba":  {"noise_scale": 0.667, "noise_scale_duration": 0.8,  "speaking_rate": 0.85},  # tonal
+    "Ewe":     {"noise_scale": 0.667, "noise_scale_duration": 0.8,  "speaking_rate": 0.85},  # tonal
+    "Hausa":   {"noise_scale": 0.5,   "noise_scale_duration": 0.6,  "speaking_rate": 0.9},
+    "Swahili": {"noise_scale": 0.5,   "noise_scale_duration": 0.6,  "speaking_rate": 0.9},
+    "Luganda": {"noise_scale": 0.5,   "noise_scale_duration": 0.6,  "speaking_rate": 0.88},
+    "Krio":    {"noise_scale": 0.4,   "noise_scale_duration": 0.3,  "speaking_rate": 0.95},  # creole, flatter prosody
+}
+
 device = "cuda" if torch.cuda.is_available() else "cpu"
 
 # ---------- Load all models & tokenizers -----------------------------------
@@ -46,17 +57,16 @@ for name, paths in TTS_MODELS.items():
     print(f"Loading {name} model...")
     model = VitsModel.from_pretrained(paths["checkpoint"]).to(device)
     model.eval()
-    # Apply clear-speech inference parameters (tweak per model if desired)
-    model.noise_scale = 0.667
-    model.noise_scale_duration = 0.7
-    model.speaking_rate = 0.85
-    models[name] = model
+    p = LANG_PARAMS[name]
+    model.noise_scale          = p["noise_scale"]
+    model.noise_scale_duration = p["noise_scale_duration"]
+    model.speaking_rate        = p["speaking_rate"]
+    models[name]     = model
     tokenizers[name] = AutoTokenizer.from_pretrained(paths["tokenizer"])
 
 # ---------- Utility: WAV ➔ MP3 Conversion -----------------------------------
 def _wav_to_mp3(wave_np: np.ndarray, sr: int) -> str:
     """Convert int16 numpy waveform to an MP3 temp file, return its path."""
-    # Ensure int16 for pydub
     if wave_np.dtype != np.int16:
         wave_np = (wave_np * 32767).astype(np.int16)
 
@@ -74,25 +84,25 @@ def tts_generate(model_name: str, text: str):
     """Generate speech for `text` using the selected model."""
     if not text:
         return None
-    model = models[model_name]
+    model     = models[model_name]
     tokenizer = tokenizers[model_name]
-    inputs = tokenizer(text, return_tensors="pt").to(device)
+    inputs    = tokenizer(text, return_tensors="pt").to(device)
     with torch.no_grad():
         wave = model(**inputs).waveform[0].cpu().numpy()
     return _wav_to_mp3(wave, model.config.sampling_rate)
 
 # ---------- Gradio Interface ------------------------------------------------
 examples = [
-    ["Yoruba", "Jídé ń ta irinṣẹ́ kápẹ́ntà àti ti ìkọ́lé ní ọjà Ìsìnkàn"],
-    ["Ewe", "amewo le atsi tre woɖo fli kple woƒe tɔkpowo kple agbawo kple galɔn wo. ʋu si nɔ tsi dram la tɔ ɖe wo xa eye nyɔnu eve yi le drɔm me le kɔkɔm ɖe tɔkpo kple galɔn me bubu hā le agba ɖe ta."],
-    ["Ewe", "ɖekakpui ene wonɔ dɔgɔe me henɔ tsi kum le teƒe aɖe to. ɖeka ɖɔ kuku se avɔ ɖe ali eye tɔkpo et̄ɔ ye nɔ wo si."],
+    ["Yoruba",  "Jídé ń ta irinṣẹ́ kápẹ́ntà àti ti ìkọ́lé ní ọjà Ìsìnkàn"],
+    ["Ewe",     "amewo le atsi tre woɖo fli kple woƒe tɔkpowo kple agbawo kple galɔn wo. ʋu si nɔ tsi dram la tɔ ɖe wo xa eye nyɔnu eve yi le drɔm me le kɔkɔm ɖe tɔkpo kple galɔn me bubu hā le agba ɖe ta."],
+    ["Ewe",     "ɖekakpui ene wonɔ dɔgɔe me henɔ tsi kum le teƒe aɖe to. ɖeka ɖɔ kuku se avɔ ɖe ali eye tɔkpo et̄ɔ ye nɔ wo si."],
     ["Swahili", "zao kusaidia kuondoa umaskini na kujenga kampeni za mwamko wa virusi vya ukimwi amezitembelea"],
     ["Swahili", "Kidole hiki ni tofauti na vidole vingine kwa sababu mwelekeo wake ni wa pekee."],
-    ["Krio", "Wetin na yu nem?"],
-    ["Krio", "aw yu de du"],
+    ["Krio",    "Wetin na yu nem?"],
+    ["Krio",    "aw yu de du"],
     ["Luganda", "Ndi musanyufu okukulaba leero"],
-    ["Hausa", "yaya za ka ƙi hafsan mafi ƙanƙanci na hafsoshin maigidana ko da yake kana dogara ga masar don kekunan yaƙi da mahayan dawakai"],
-    ["Hausa", "ina fata dukkanku za ku ji ni sosai. wannan ita ce ma'anar kawai."]
+    ["Hausa",   "yaya za ka ƙi hafsan mafi ƙanƙanci na hafsoshin maigidana ko da yake kana dogara ga masar don kekunan yaƙi da mahayan dawakai"],
+    ["Hausa",   "ina fata dukkanku za ku ji ni sosai. wannan ita ce ma'anar kawai."]
 ]
 
 demo = gr.Interface(
@@ -103,12 +113,10 @@ demo = gr.Interface(
     ],
     outputs=gr.Audio(type="filepath", label="Audio", autoplay=True),
     title="Multi‐Model Text-to-Speech",
-    description=(
-        "Select a TTS model from the dropdown and enter text to generate speech."
-    ),
+    description="Select a TTS model from the dropdown and enter text to generate speech.",
     examples=examples,
     cache_examples=True,
 )
 
 if __name__ == "__main__":
-    demo.launch()
+    demo.launch()