Upload app.py with huggingface_hub

app.py

@@ -1,16 +1,14 @@
 
-import gradio as gr
 import torch, yaml, os, sys, glob, re
 import librosa
 import soundfile as sf
 import torchaudio
+import numpy as np
+from fastapi import FastAPI
+from fastapi.responses import FileResponse
+from pydantic import BaseModel
 from huggingface_hub import snapshot_download
-from munch import Munch
-from nltk.tokenize import word_tokenize
-import nltk
-nltk.download("punkt_tab", quiet=True)
 
-# ── Clone + patch StyleTTS2 ────────────────────────────────
 os.system("git clone https://github.com/yl4579/StyleTTS2 /app/StyleTTS2 2>/dev/null || true")
 
 for fpath in ["/app/StyleTTS2/models.py", "/app/StyleTTS2/utils.py"]:
@@ -24,7 +22,6 @@ for fpath in ["/app/StyleTTS2/models.py", "/app/StyleTTS2/utils.py"]:
     )
     with open(fpath, "w") as f:
         f.write(patched)
-print("Patched torch.load")
 
 sys.path.insert(0, "/app/StyleTTS2")
 os.chdir("/app/StyleTTS2")
@@ -35,6 +32,10 @@ from models import *
 from utils import *
 from text_utils import TextCleaner
 from Modules.diffusion.sampler import DiffusionSampler, ADPM2Sampler, KarrasSchedule
+from munch import Munch
+from nltk.tokenize import word_tokenize
+import nltk
+nltk.download("punkt_tab", quiet=True)
 
 device      = "cpu"
 textcleaner = TextCleaner()
@@ -107,17 +108,30 @@ ref_candidates = (
     glob.glob(f"{model_dir}/ref_wavs/*.wav") +
     glob.glob(f"{model_dir}/*.wav")
 )
-DEFAULT_REF   = sorted(ref_candidates)[0]
-DEFAULT_STYLE = compute_style(DEFAULT_REF)
-print(f"Reference: {DEFAULT_REF}")
+DEFAULT_STYLE = compute_style(sorted(ref_candidates)[0])
+print("Ready")
+
+# ── FastAPI ────────────────────────────────────────────────
+app = FastAPI()
 
-def synthesize(text, ref_audio=None, alpha=0.3, beta=0.7, steps=5):
+class TTSRequest(BaseModel):
+    text: str
+    alpha: float = 0.3
+    beta:  float = 0.7
+    steps: int   = 5
+
+@app.get("/")
+def root():
+    return {"status": "ok", "model": "luhya-tts"}
+
+@app.post("/predict")
+def predict(req: TTSRequest):
     import phonemizer
     pb    = phonemizer.backend.EspeakBackend(
         language="sw", preserve_punctuation=True, with_stress=True)
-    ref_s = compute_style(ref_audio) if ref_audio else DEFAULT_STYLE
+    ref_s = DEFAULT_STYLE
 
-    ps     = " ".join(word_tokenize(pb.phonemize([text.strip()])[0]))
+    ps     = " ".join(word_tokenize(pb.phonemize([req.text.strip()])[0]))
     tokens = textcleaner(ps)
     tokens.insert(0, 0)
     tokens = torch.LongTensor(tokens).to(device).unsqueeze(0)
@@ -132,11 +146,11 @@ def synthesize(text, ref_audio=None, alpha=0.3, beta=0.7, steps=5):
         sp = sampler(
             noise=torch.randn((1, 256)).unsqueeze(1).to(device),
             embedding=bd, embedding_scale=1,
-            features=ref_s, num_steps=int(steps)
+            features=ref_s, num_steps=req.steps
         ).squeeze(1)
 
-        s   = beta  * sp[:, 128:] + (1 - beta)  * ref_s[:, 128:]
-        ref = alpha * sp[:, :128] + (1 - alpha) * ref_s[:, :128]
+        s   = req.beta  * sp[:, 128:] + (1 - req.beta)  * ref_s[:, 128:]
+        ref = req.alpha * sp[:, :128] + (1 - req.alpha) * ref_s[:, :128]
 
         d    = model.predictor.text_encoder(d_en, s, il, tm)
         x, _ = model.predictor.lstm(d)
@@ -166,19 +180,8 @@ def synthesize(text, ref_audio=None, alpha=0.3, beta=0.7, steps=5):
 
     wav = out.squeeze().cpu().numpy()[..., :-50]
     sf.write("/tmp/output.wav", wav, 24000)
-    return "/tmp/output.wav"
-
-demo = gr.Interface(
-    fn      = synthesize,
-    inputs  = [
-        gr.Textbox(label="Luhya text", value="mirembe. obulani lwa bwana nyasaye.", lines=3),
-        gr.Audio(label="Reference voice (optional)", type="filepath", value=None),
-        gr.Slider(0.0, 1.0, value=0.3, step=0.1, label="Alpha"),
-        gr.Slider(0.0, 1.0, value=0.7, step=0.1, label="Beta"),
-        gr.Slider(1, 10, value=5, step=1, label="Diffusion steps"),
-    ],
-    outputs = gr.Audio(label="Generated speech", type="filepath"),
-    title   = "Luhya (Lunyore) TTS",
-)
+    return FileResponse("/tmp/output.wav", media_type="audio/wav")
 
-demo.launch(server_name="0.0.0.0", server_port=7860, share=False, inline=False)
+if __name__ == "__main__":
+    import uvicorn
+    uvicorn.run(app, host="0.0.0.0", port=7860)