Default back to model pauses, speed 1.0 (listening test); clause mode experimental

infer.py

@@ -7,9 +7,9 @@ internal-silence cap (0.6 s), 0.6 s between sentences.
 
   python infer.py --text "<eng>this is dynamic accent,</eng> <pcm>e dey sound nice abi?</pcm>" --seed 0 --out out.wav
 
-Prosody v2 (default, `--pause_mode clause`): the text is cut at , ; : . ? ! and at language-switch points; each chunk
+Prosody v2 (experimental, `--pause_mode clause`; default is `model`): the text is cut at , ; : . ? ! and at language-switch points; each chunk
 is generated with the same LM/flow/prompt, its edge silence is trimmed (energy VAD, 30 ms kept), its speech is sped
-up pitch-preserving by `--speed` (WSOLA, default 1.05), and chunks are joined with set gaps from the real anchor's pause statistics:
+up pitch-preserving by `--speed` (WSOLA, default 1.0), and chunks are joined with set gaps from the real anchor's pause statistics:
 comma/;/: 0.40 s (+-0.05 jitter), sentence end 0.60 s, unpunctuated language switch 0.12 s, with 10 ms fades.
 `--pause_mode model --speed 1.0` reproduces the Phase 2 final samples (model-generated pauses, sentence stitching).
 
@@ -27,7 +27,7 @@ HERE = os.path.dirname(os.path.abspath(__file__))
 LANGS = ["eng", "ibo", "hau", "yor", "pcm"]
 L2I = {l: i for i, l in enumerate(LANGS)}
 ANCHOR = "waxal_yor_4"
-DEFAULT_SPEED = 1.05   # from the speaking-rate measurement (README, "Prosody v2"): median ref/system ratio 1.06
+DEFAULT_SPEED = 1.0    # user listening test 2026-10-21: original model pauses preferred; clause/1.05 kept as experimental
 
 
 # ------------------------------------------------------------------ text front-end (tags -> spans)
@@ -343,7 +343,7 @@ class DynamicAccent:
             sp = sp + [0] * 3
         return token2wav(self.cosy, sp, *P["flow"])
 
-    def say(self, tagged_text, seed=0, prompt_k=0, stitch=True, pause_mode="clause", speed=1.10, return_info=False):
+    def say(self, tagged_text, seed=0, prompt_k=0, stitch=True, pause_mode="model", speed=1.0, return_info=False):
         if pause_mode == "clause":
             return self.say_clause(tagged_text, seed, prompt_k, speed, return_info)
         w = self.say_model(tagged_text, seed, prompt_k, stitch)
@@ -400,7 +400,7 @@ def main():
     ap.add_argument("--dtype", default="auto", choices=["auto", "bf16", "fp16"],
                     help="LM autocast dtype. Samples were made with bf16; T4 has no bf16, so auto picks fp16 there")
     ap.add_argument("--no_stitch", action="store_true", help="generate multi-sentence text in one pass")
-    ap.add_argument("--pause_mode", default="clause", choices=["clause", "model"],
+    ap.add_argument("--pause_mode", default="model", choices=["clause", "model"],
                     help="clause: clause-level chunks joined with anchor-matched gaps (default); model: Phase 2 final behaviour")
     ap.add_argument("--speed", type=float, default=DEFAULT_SPEED,
                     help="pitch-preserving tempo factor applied to speech (gaps are kept as set). 1.0 = Phase 2 final pace")