Update app.py

app.py

@@ -3,64 +3,72 @@ import os
 import tempfile
 from datetime import datetime
 
-import librosa
 import numpy as np
 import soundfile as sf
 import streamlit as st
 from TTS.api import TTS
 
-st.set_page_config(page_title="Urdu Voice Cloner", page_icon="🗣️", layout="centered")
+st.set_page_config(page_title="Urdu Voice Cloner (XTTS v2)", page_icon="🗣️", layout="centered")
 
 st.title("🗣️ Urdu Text → Your Voice (Voice Cloning)")
-st.caption("Upload a short sample of your voice, type Urdu text, and get audio in your voice.")
+st.caption("Upload a short sample of your voice, type Urdu text, and get audio in your voice (XTTS v2, CPU friendly).")
 
 # ----------------------------
-# Caching the model to avoid reloading
+# Cache the model so it loads once
 # ----------------------------
 @st.cache_resource(show_spinner=True)
 def load_tts():
-    # XTTS v2 supports multilingual zero-shot cloning, including Urdu (code: 'ur')
-    # Model will download on first run and then be cached by the Space
+    # Multilingual zero-shot cloning, supports Urdu with language='ur'
     return TTS(model_name="tts_models/multilingual/multi-dataset/xtts_v2")
 
 tts = load_tts()
 
 # ----------------------------
-# Sidebar: options
+# Sidebar options
 # ----------------------------
 with st.sidebar:
     st.header("Options")
-    st.markdown("**Reference voice**")
-    st.caption("Upload a clean 10–30 second sample with minimal noise.")
-    # XTTS controls
+    st.caption("Upload a clean 10–30s clip, no background noise if possible.")
     similarity_boost = st.slider("Similarity boost", 0.0, 1.0, 0.75, 0.05)
-    stability = st.slider("Stability", 0.0, 1.0, 0.6, 0.05)
+    stability = st.slider("Stability", 0.0, 1.0, 0.60, 0.05)
     style = st.slider("Style (expressiveness)", 0.0, 1.0, 0.35, 0.05)
-    seed = st.number_input("Random seed (for reproducibility)", value=42, step=1)
-
-    st.markdown("---")
-    st.markdown("**Post-processing**")
-    rate = st.slider("Speaking rate (time-stretch)", 0.75, 1.25, 1.00, 0.01)
     normalize = st.checkbox("Normalize loudness", True)
-
-    st.markdown("---")
     base_name = st.text_input("Output filename (no extension)", "urdu_voice_clone")
+    seed = st.number_input("Random seed", value=42, step=1)
 
 # ----------------------------
-# Inputs
+# Simple helpers (no librosa)
 # ----------------------------
-ref_file = st.file_uploader(
-    "Upload your voice sample (wav/mp3/m4a)", 
-    type=["wav", "mp3", "m4a", "ogg", "flac"]
-)
+def simple_trim_silence(wave: np.ndarray, threshold: float = 1e-4, pad: int = 0) -> np.ndarray:
+    """
+    Very simple silence trim: finds where absolute amplitude exceeds threshold.
+    If nothing exceeds threshold, returns original.
+    """
+    if wave.ndim > 1:
+        wave = wave.mean(axis=1)
+    idx = np.where(np.abs(wave) > threshold)[0]
+    if idx.size == 0:
+        return wave
+    start = max(int(idx[0]) - pad, 0)
+    end = min(int(idx[-1]) + pad, wave.shape[0])
+    return wave[start:end]
+
+def normalize_peak(wave: np.ndarray, peak: float = 0.98) -> np.ndarray:
+    m = np.max(np.abs(wave)) + 1e-9
+    return (peak * wave / m).astype(np.float32)
+
+def wav_bytes_from_array(y: np.ndarray, sr: int) -> bytes:
+    buf = io.BytesIO()
+    sf.write(buf, y, sr, format="WAV")
+    buf.seek(0)
+    return buf.read()
 
+# ----------------------------
+# Inputs
+# ----------------------------
+ref_file = st.file_uploader("Upload your voice sample (wav/mp3/m4a/ogg/flac)", type=["wav", "mp3", "m4a", "ogg", "flac"])
 default_text = "یہ میری آواز کی مثال ہے۔ آپ یہاں اپنا متن لکھیں اور آڈیو حاصل کریں۔"
-text = st.text_area(
-    "Urdu text",
-    value=default_text,
-    height=180,
-    placeholder="یہاں اردو میں ٹیکسٹ لکھیں یا پیسٹ کریں…"
-)
+text = st.text_area("Urdu text", value=default_text, height=180, placeholder="یہاں اردو میں ٹیکسٹ لکھیں یا پیسٹ کریں…")
 
 col1, col2 = st.columns(2)
 with col1:
@@ -70,45 +78,10 @@ with col2:
 
 if clear_btn:
     st.session_state.pop("audio_bytes", None)
-    st.session_state.pop("preview_sr", None)
     st.experimental_rerun()
 
 # ----------------------------
-# Helpers
-# ----------------------------
-def load_and_standardize(audio_file, target_sr=16000):
-    """Load user audio, convert to mono 16 kHz WAV bytes and return temp path."""
-    y, sr = librosa.load(audio_file, sr=None, mono=True)
-    if len(y) < target_sr * 3:
-        st.warning("Voice sample is very short. Try at least 5–10 seconds for better cloning.")
-    y_res = librosa.resample(y, orig_sr=sr, target_sr=target_sr)
-    # Light trim to remove leading/trailing silence
-    yt, _ = librosa.effects.trim(y_res, top_db=30)
-    if yt.size < target_sr:  # ensure at least 1s remains
-        yt = y_res
-    tmp_wav = tempfile.NamedTemporaryFile(delete=False, suffix=".wav")
-    sf.write(tmp_wav.name, yt, target_sr)
-    return tmp_wav.name
-
-def postprocess_rate_and_norm(wav, sr, rate_factor=1.0, do_norm=True):
-    """Time-stretch and normalize loudness."""
-    y = wav.astype(np.float32)
-    if rate_factor != 1.0:
-        # librosa requires strictly positive values
-        y = librosa.effects.time_stretch(y, rate_factor)
-    if do_norm:
-        peak = np.max(np.abs(y)) + 1e-9
-        y = 0.98 * (y / peak)
-    return y
-
-def wav_bytes_from_array(y, sr):
-    buf = io.BytesIO()
-    sf.write(buf, y, sr, format="WAV")
-    buf.seek(0)
-    return buf.read()
-
-# ----------------------------
-# Run
+# Run synthesis
 # ----------------------------
 if run_btn:
     if not text.strip():
@@ -117,76 +90,72 @@ if run_btn:
         st.warning("براہ کرم اپنی آواز کی آڈیو فائل اپلوڈ کریں۔")
     else:
         try:
-            st.info("Preparing reference voice…")
-            ref_path = load_and_standardize(ref_file)
+            # Save uploaded file to a temp path (XTTS can accept various formats via soundfile/ffmpeg backend)
+            tmp_ref = tempfile.NamedTemporaryFile(delete=False, suffix=f".{ref_file.name.split('.')[-1]}")
+            tmp_ref.write(ref_file.read())
+            tmp_ref.flush()
+            tmp_ref.close()
+
+            # Optional: quick silence trim to reduce leading/trailing gaps
+            try:
+                y_ref, sr_ref = sf.read(tmp_ref.name, dtype="float32", always_2d=False)
+                y_ref = simple_trim_silence(y_ref)
+                sf.write(tmp_ref.name, y_ref, sr_ref)  # overwrite trimmed
+            except Exception:
+                # If reading/trim fails, keep original file
+                pass
+
+            st.info("Cloning voice and synthesizing Urdu… (CPU can take a bit on first run)")
 
-            st.info("Cloning voice and synthesizing Urdu…")
-            # Generate to a temporary file first
             out_wav_path = tempfile.NamedTemporaryFile(delete=False, suffix=".wav").name
 
-            # Coqui XTTS generation
-            # Extra params passed via "speaker_wav" and "language"
-            # Controls: "speaker_similarity", "style", "temperature", "length_scale" etc. are model dependent.
+            # Generate audio
             tts.tts_to_file(
                 text=text.strip(),
                 file_path=out_wav_path,
-                speaker_wav=ref_path,
+                speaker_wav=tmp_ref.name,
                 language="ur",
-                # Extra inference kwargs routed to the model (supported by XTTS v2)
-                # See: https://github.com/coqui-ai/TTS
-                # Using similarity/stability/style through speaker conditioning
-                # Some builds accept these as speaker_cfg; we forward common names:
+                # Common conditioning knobs
+                speaker_similarity=float(similarity_boost),
+                stability=float(stability),
+                style=float(style),
                 split_sentences=True,
-                speed=1.0,  # base speed, we will also post-process rate
-                speaker_similarity=similarity_boost,
-                stability=stability,
-                style_wav=None,
-                style=style,
-                seed=int(seed)
+                seed=int(seed),
             )
 
-            # Read back and post-process
-            y, sr = sf.read(out_wav_path, dtype="float32")
-            y = postprocess_rate_and_norm(y, sr, rate_factor=rate, do_norm=normalize)
-            audio_bytes = wav_bytes_from_array(y, sr)
+            # Load, optional normalize, then serve
+            y, sr = sf.read(out_wav_path, dtype="float32", always_2d=False)
+            if normalize:
+                y = normalize_peak(y)
 
-            # Stash in session for preview and download
+            audio_bytes = wav_bytes_from_array(y, sr)
             st.session_state["audio_bytes"] = audio_bytes
-            st.session_state["preview_sr"] = sr
 
-            # Clean temp files
+            # Cleanup temp files
             try:
-                os.remove(ref_path)
+                os.remove(tmp_ref.name)
                 os.remove(out_wav_path)
             except Exception:
                 pass
 
-            st.success("آڈیو تیار ہے۔ نیچے سنیں یا ڈاؤن لوڈ کریں۔")
+            st.success("آڈیو تیار ہے۔")
 
         except Exception as e:
             st.error(f"کچھ مسئلہ آیا: {e}")
 
 # ----------------------------
-# Preview and download
+# Preview & download
 # ----------------------------
 if "audio_bytes" in st.session_state:
     st.markdown("### ▶️ Preview")
     st.audio(st.session_state["audio_bytes"], format="audio/wav")
 
     ts = datetime.now().strftime("%Y%m%d_%H%M%S")
-    base = (base_name or "urdu_voice_clone").strip()
-    fname = f"{base}_{ts}.wav"
-
-    st.download_button(
-        "⬇️ Download WAV",
-        data=st.session_state["audio_bytes"],
-        file_name=fname,
-        mime="audio/wav",
-        use_container_width=True
-    )
+    fname = f"{(base_name or 'urdu_voice_clone').strip()}_{ts}.wav"
+    st.download_button("⬇️ Download WAV", data=st.session_state["audio_bytes"], file_name=fname, mime="audio/wav", use_container_width=True)
 
 st.markdown("---")
 st.caption(
-    "Tips: Use a clear 10–30 second reference with low noise. Speak naturally. "
-    "If cloning feels off, try a different sample, raise Similarity, or lower Stability a little."
+    "Tips: Use a clear 10–30 second reference with low noise. If cloning feels off, try a different sample, "
+    "raise Similarity slightly, or lower Stability a little."
 )