perf: reuse GCC-PHAT TDOA; skip 16 kHz AudioNormalizer

Delay-and-Sum still uses the tutorial STFT/cov/GCC-PHAT/DelaySum/ISTFT
path. TDOA is refreshed every BEAMFORM_TDOA_EVERY hops (default 5).

audio_cleanup.py

@@ -64,7 +64,8 @@ class AudioCleaner:
                  chunk_s: float = 1.0,
                  hop_s: float = 0.2,
                  beamform: str = "auto",
-                 torch_num_threads: int = 1):
+                 torch_num_threads: int = 1,
+                 tdoa_every: int = 5):
         self.mode = mode
         self.beamform_mode = beamform
         self.sample_rate = sample_rate
@@ -78,6 +79,14 @@ class AudioCleaner:
         self.backend = "off"
         self.last_dt_ms = -1.0
         self.torch_num_threads = max(1, int(torch_num_threads))
+        self.tdoa_every = max(1, int(tdoa_every))
+        self._stft = None
+        self._cov = None
+        self._gccphat = None
+        self._delaysum = None
+        self._istft = None
+        self._last_tdoas = None
+        self._tdoa_age = 0
 
         self._load_torch()
         if beamform == "force":
@@ -130,12 +139,23 @@ class AudioCleaner:
             return
         try:
             log.info("loading speechbrain DelaySum_Beamformer (GCC-PHAT)")
-            self._normalizer = AudioNormalizer(
-                sample_rate=self.sample_rate, mix="keep")
+            # Native 16 kHz C920s: AudioNormalizer would resample 16k->16k.
+            if self.sample_rate != SAMPLE_RATE:
+                self._normalizer = AudioNormalizer(
+                    sample_rate=self.sample_rate, mix="keep")
+            else:
+                self._normalizer = None
             bf = DelaySum_Beamformer(sampling_rate=self.sample_rate)
             bf.eval()
             self._beamformer = bf
-            log.info("DelaySum beamformer ready (stereo in, mono out)")
+            self._stft = bf.stft
+            self._cov = bf.cov
+            self._gccphat = bf.gccphat
+            self._delaysum = bf.delaysum
+            self._istft = bf.istft
+            log.info(
+                "DelaySum beamformer ready (stereo in, mono out, tdoa_every=%d)",
+                self.tdoa_every)
         except Exception as e:  # noqa: BLE001
             log.warning("failed to load DelaySum_Beamformer: %s", e)
             self._beamformer = None
@@ -216,9 +236,18 @@ class AudioCleaner:
         wav = torch.from_numpy(np.ascontiguousarray(stereo, dtype=np.float32))
         if self._normalizer is not None:
             wav = self._normalizer(wav, self.sample_rate)
-        # DelaySum_Beamformer expects [batch, time, channels]
+        # DelaySum expects [batch, time, channels]
+        xs = wav.unsqueeze(0)
         with torch.no_grad():
-            y = self._beamformer(wav.unsqueeze(0))
+            Xs = self._stft(xs)
+            if (self._last_tdoas is None
+                    or self._tdoa_age >= self.tdoa_every):
+                XXs = self._cov(Xs)
+                self._last_tdoas = self._gccphat(XXs)
+                self._tdoa_age = 0
+            self._tdoa_age += 1
+            Ys = self._delaysum(Xs, self._last_tdoas)
+            y = self._istft(Ys)
         out = y.squeeze().detach().cpu().numpy().astype(np.float32, copy=False)
         if out.ndim > 1:
             out = np.reshape(out, -1)