Upload app.py

app.py

@@ -565,17 +565,20 @@ def _take_score(i, med):
     return s
 
 def synth_best(gen_fn, text, attempts=1, asr_fn=None, sr=XTTS_SR, seed=1234,
-               accept_sim=0.90, secs_fn=None, lang="ar", guard_retries=2):
+               accept_sim=0.90, secs_fn=None, lang="ar", guard_retries=2, min_secs=0.45):
     """gen_fn(seed:int) → float32 wav.  Returns (wav, info, n_generated).
     attempts==1 : cheap self-healing guard — regenerate (≤ guard_retries) only if the
-                  take is obviously broken (silent / absurd speech-rate / clipped).
-    attempts>1  : best-of-N ranked by ASR similarity + speaker similarity + duration."""
+                  take is obviously broken (silent / absurd speech-rate / clipped / off-speaker).
+    attempts>1  : best-of-N ranked by ASR similarity + speaker similarity + duration.
+    `secs_fn` (speaker-encoder cosine similarity) is a single extra forward pass — cheap
+    enough to run even when attempts==1, and it is the only check that catches a take
+    that drifted away from the reference speaker's timbre (silence/clip/rate checks don't)."""
     multi = attempts > 1
     takes, k = [], 0
     while True:
         wav = trim_take(gen_fn(seed + 7919 * k), sr)
         info = take_info(wav, sr, text, asr_fn if multi else None, lang)
-        if multi and secs_fn is not None and len(wav) > int(0.3 * sr):
+        if secs_fn is not None and len(wav) > int(0.3 * sr):
             try: info["secs"] = float(secs_fn(wav, sr))
             except Exception: info["secs"] = None
         takes.append((wav, info)); k += 1
@@ -583,7 +586,8 @@ def synth_best(gen_fn, text, attempts=1, asr_fn=None, sr=XTTS_SR, seed=1234,
             if k >= attempts: break
             if info["ok"] and info["sim"] is not None and info["sim"] >= accept_sim: break
         else:
-            if (info["ok"] and info["clip"] < 0.005) or k > guard_retries: break
+            secs_ok = info.get("secs") is None or info["secs"] >= min_secs
+            if (info["ok"] and info["clip"] < 0.005 and secs_ok) or k > guard_retries: break
     if len(takes) == 1: return takes[0][0], takes[0][1], 1
     med = float(np.median([i["dur"] for _, i in takes]))
     best = max(takes, key=lambda t: _take_score(t[1], med))
@@ -1938,7 +1942,12 @@ def _run_tts_quality(model, text, refs, lang, base_params, out_path, job_id,
 
     d        = _model_defaults(model)
     try:
-        secs_fn = _secs_fn(model, lat_spk) if attempts > 1 else None
+        # One speaker-encoder forward pass per take — cheap, so run it even at attempts==1;
+        # this used to be gated behind attempts>1, which meant the default UI setting
+        # (attempts=1) never actually checked whether a chunk still sounded like the
+        # reference speaker. That gap is the main source of "sounds like a different
+        # speaker sometimes" on normal (non best-of-N) generations.
+        secs_fn = _secs_fn(model, lat_spk)
     except Exception as ex:                      # scoring is a bonus — never block generation
         print(f"[!] speaker-similarity scoring disabled: {ex}", flush=True)
         secs_fn = None
@@ -1968,10 +1977,16 @@ def _run_tts_quality(model, text, refs, lang, base_params, out_path, job_id,
                                        secs_fn=secs_fn, lang=lang)
         retries += _n - 1
         wav = _residual_stretch(wav, resid)
+        # `conf` (how confident analyze_phrase_emotion was) was computed above but never
+        # reached these two effects — they used to apply full-strength dynamics/pause
+        # changes even on a low-confidence (often false-positive) keyword match, which is
+        # what made ordinary neutral sentences suddenly jump in loudness/pacing next to
+        # their neighbours. Scaling by `conf` makes weak matches barely audible and keeps
+        # strong matches (question marks, "!!", clear keywords) at full strength.
         if use_var and emo != "neutral":
-            wav = apply_phrase_dynamics(wav, vq.XTTS_SR, emo, variation * 0.5)
+            wav = apply_phrase_dynamics(wav, vq.XTTS_SR, emo, variation * 0.5 * conf)
         gap = EMOTION_DELTAS.get(emo, EMOTION_DELTAS["neutral"])[3]
-        emo_factor = 1.0 + (gap / 110.0 - 1.0) * 0.5 * min(variation, 1.0) if use_var else 1.0
+        emo_factor = 1.0 + (gap / 110.0 - 1.0) * 0.5 * min(variation, 1.0) * conf if use_var else 1.0
         takes.append(wav); pauses.append(pz * scale * emo_factor)
         job_progress[job_id]["current"] = i + 1