Update denoiser.py

denoiser.py

@@ -1,20 +1,17 @@
 """
 Denoiser — Cleanvoice SDK (cleanvoice-sdk)
-No Cloudinary needed — SDK supports local file path directly.
 
-QUALITY FIX:
-  - Cleanvoice receives the ORIGINAL quality WAV (48kHz stereo preserved)
-  - 16kHz mono conversion is only done for Whisper transcription (in transcriber.py)
-  - Final output preserves full original sample rate + channels
+FIX: Send the RAW original audio directly to Cleanvoice API — no conversion,
+     no touching, no format changes. Cleanvoice accepts all common formats.
+     16kHz mono conversion is ONLY done in transcriber.py for Whisper.
 
 Install: pip install cleanvoice-sdk
-Only requires: CLEANVOICE_API_KEY in HF Space secrets.
+Requires: CLEANVOICE_API_KEY in HF Space secrets.
 """
 
 import os
 import re
 import logging
-import subprocess
 
 logger = logging.getLogger(__name__)
 
@@ -23,14 +20,14 @@ CLEANVOICE_API_KEY = os.environ.get(
     "cvk_r1z_P4IyEBE.Oy3JnO2_WS3msTTzYhccTLqzcE1Tp2wZgI02vHF17BA"
 )
 
-# Filler words removed from transcript text (text-side cleanup)
+# Filler words removed from transcript text (text-side cleanup only)
 FILLER_PATTERN = r'\b(um|uh|hmm|basically|like|ante|ane|matlab|yani|acha|right|okay)\b'
 
 
 class Denoiser:
     def __init__(self):
         if CLEANVOICE_API_KEY:
-            print("[Denoiser] ✅ Cleanvoice SDK ready (no Cloudinary needed)")
+            print("[Denoiser] ✅ Cleanvoice SDK ready")
         else:
             print("[Denoiser] ⚠️  No CLEANVOICE_API_KEY — denoising will be skipped")
 
@@ -43,67 +40,37 @@ class Denoiser:
                 remove_silences: bool = True,
                 **kwargs) -> dict:
         """
-        Process audio using cleanvoice-sdk.
-        QUALITY FIX: Sends original quality to Cleanvoice (NOT downsampled).
+        Send raw audio directly to Cleanvoice — no pre-processing.
+        Cleanvoice handles MP3, WAV, AAC, FLAC, OGG, OPUS, M4A natively.
         Returns {'audio_path': str, 'stats': dict}
         """
         stats = {"noise_method": "Cleanvoice SDK"}
 
-        # ── Step 1: Convert to WAV but PRESERVE original sample rate + channels ──
-        # FIX: Do NOT force 16kHz mono here — that was destroying quality.
-        # 16kHz mono is only needed for Whisper (handled in transcriber.py).
-        # Cleanvoice works best with original quality audio.
-        standard_wav = os.path.join(out_dir, "input_converted.wav")
-
-        # Get original sample rate and channels first
-        try:
-            import subprocess as sp
-            probe = sp.run([
-                "ffprobe", "-v", "error",
-                "-show_entries", "stream=sample_rate,channels",
-                "-of", "default=noprint_wrappers=1:nokey=1",
-                audio_path
-            ], capture_output=True, text=True)
-            lines = probe.stdout.strip().split("\n")
-            orig_sr  = int(lines[0]) if lines[0].isdigit() else 48000
-            orig_ch  = int(lines[1]) if len(lines) > 1 and lines[1].isdigit() else 2
-        except Exception:
-            orig_sr, orig_ch = 48000, 2
-
-        print(f"[Denoiser] Original audio: {orig_sr}Hz, {orig_ch}ch")
-
-        # Convert to WAV keeping original sample rate and channels
-        result = subprocess.run([
-            "ffmpeg", "-y", "-i", audio_path,
-            "-ar", str(orig_sr),   # preserve original sample rate
-            "-ac", str(orig_ch),   # preserve original channels
-            "-acodec", "pcm_s16le",
-            standard_wav
-        ], capture_output=True)
-
-        if result.returncode != 0:
-            logger.error(f"[Denoiser] FFmpeg conversion failed: {result.stderr.decode()}")
-            return {"audio_path": audio_path, "stats": stats}
-
-        print(f"[Denoiser] ✅ Converted to WAV ({orig_sr}Hz, {orig_ch}ch) — quality preserved")
-
-        # ── Step 2: Check API key ─────────────────────────────────────
+        # ── Check API key ─────────────────────────────────────────────
         if not CLEANVOICE_API_KEY:
-            logger.warning("[Denoiser] No API key — skipping, returning converted WAV")
+            logger.warning("[Denoiser] No API key — skipping denoising")
             stats["noise_method"] = "Skipped (no API key)"
-            return {"audio_path": standard_wav, "stats": stats}
+            return {"audio_path": audio_path, "stats": stats}
 
-        # ── Step 3: Run Cleanvoice SDK ────────────────────────────────
+        # ── Send RAW audio directly to Cleanvoice ─────────────────────
+        # FIX: Do NOT convert or touch the audio before sending.
+        # The old code was converting to 16kHz mono WAV first,
+        # which degraded quality and likely confused Cleanvoice.
+        # Cleanvoice accepts the original file as-is.
         try:
             from cleanvoice import Cleanvoice
 
-            client    = Cleanvoice.from_env()
-            clean_wav = os.path.join(out_dir, "cleanvoice_output.wav")
+            client = Cleanvoice(api_key=CLEANVOICE_API_KEY)
+
+            # Determine output filename (keep same extension as input)
+            ext = os.path.splitext(audio_path)[1] or ".wav"
+            clean_out = os.path.join(out_dir, f"cleanvoice_output{ext}")
 
-            print(f"[Denoiser] Sending original quality audio to Cleanvoice...")
+            print(f"[Denoiser] Sending raw audio to Cleanvoice: {audio_path}")
+            print(f"[Denoiser] File size: {os.path.getsize(audio_path) / 1024:.1f} KB")
 
             result_cv = client.process(
-                standard_wav,        # original quality WAV ✅
+                audio_path,              # RAW original file — no conversion ✅
                 remove_noise=True,
                 studio_sound=True,
                 normalize=True,
@@ -115,22 +82,26 @@ class Denoiser:
                 breath=True,
             )
 
-            result_cv.audio.download(clean_wav)
-            print(f"[Denoiser] ✅ Cleanvoice done — saved to {clean_wav}")
-
-            stats.update({
-                "noise_method":     "Cleanvoice SDK ✅",
-                "fillers_removed":  "yes" if remove_fillers  else "no",
-                "stutters_removed": "yes" if remove_stutters else "no",
-                "silences_removed": "yes" if remove_silences else "no",
-            })
-            return {"audio_path": clean_wav, "stats": stats}
+            result_cv.audio.download(clean_out)
+
+            if os.path.exists(clean_out) and os.path.getsize(clean_out) > 0:
+                print(f"[Denoiser] ✅ Cleanvoice done → {clean_out} "
+                      f"({os.path.getsize(clean_out) / 1024:.1f} KB)")
+                stats.update({
+                    "noise_method":     "Cleanvoice SDK ✅",
+                    "fillers_removed":  "yes" if remove_fillers  else "no",
+                    "stutters_removed": "yes" if remove_stutters else "no",
+                    "silences_removed": "yes" if remove_silences else "no",
+                })
+                return {"audio_path": clean_out, "stats": stats}
+            else:
+                raise RuntimeError("Cleanvoice returned empty/missing file")
 
         except Exception as e:
-            logger.error(f"[Denoiser] Cleanvoice SDK failed: {e}", exc_info=True)
-            print(f"[Denoiser] ⚠️ Cleanvoice failed ({e}) — returning converted WAV")
+            logger.error(f"[Denoiser] Cleanvoice failed: {e}", exc_info=True)
+            print(f"[Denoiser] ⚠️ Cleanvoice failed ({e}) — returning original audio unchanged")
             stats["noise_method"] = f"Cleanvoice failed: {str(e)[:120]}"
-            return {"audio_path": standard_wav, "stats": stats}
+            return {"audio_path": audio_path, "stats": stats}
 
     # ══════════════════════════════════════════════════════════════════
     # TEXT-SIDE FILLER CLEANUP