From 65a0ef67cf027b32b798eaf8af3700d9dc25da80 Mon Sep 17 00:00:00 2001
From: Vuong Hoang <vh@phasefinal.com>
Date: Mon, 1 Jun 2026 23:52:05 -0700
Subject: [PATCH] fix(chatterbox-fast): correct-length WAV header for one-shot
 responses

stream=false + format=wav emitted the streaming 0xFFFFFFFF-length header, so a
buffered consumer reading a complete wav got bogus RIFF/data sizes. One-shot knows
the full length, so emit correct sizes; streaming keeps the open-ended header
(length genuinely unknown up front). Verified remote: one-shot wav data size ==
bytes-44, python wave.open() reads 2.20s cleanly; streaming still 0xFFFFFFFF.
---
 stacks/chatterbox-fast/app.py | 25 +++++++++++++++++--------
 1 file changed, 17 insertions(+), 8 deletions(-)

diff --git a/stacks/chatterbox-fast/app.py b/stacks/chatterbox-fast/app.py
index 91f79d4..758456b 100644
--- a/stacks/chatterbox-fast/app.py
+++ b/stacks/chatterbox-fast/app.py
@@ -217,12 +217,15 @@ def _pcm16(wav: torch.Tensor) -> bytes:
     return (a * 32767.0).astype("<i2").tobytes()
 
 
-def _wav_header(sr: int) -> bytes:
-    """Streaming WAV header with unknown length (0xFFFFFFFF sizes)."""
+def _wav_header(sr: int, data_len: int | None = None) -> bytes:
+    """WAV header. data_len=None → streaming (length unknown, 0xFFFFFFFF sizes,
+    player reads to EOF); an int → correct RIFF/data sizes for a complete file."""
+    data_size = 0xFFFFFFFF if data_len is None else data_len
+    riff_size = 0xFFFFFFFF if data_len is None else 36 + data_len
     return (
-        b"RIFF" + struct.pack("<I", 0xFFFFFFFF) + b"WAVE"
+        b"RIFF" + struct.pack("<I", riff_size) + b"WAVE"
         + b"fmt " + struct.pack("<IHHIIHH", 16, 1, 1, sr, sr * 2, 2, 16)
-        + b"data" + struct.pack("<I", 0xFFFFFFFF)
+        + b"data" + struct.pack("<I", data_size)
     )
 
 
@@ -276,19 +279,25 @@ def tts(req: TTSRequest) -> StreamingResponse:
         # workload); concurrent callers queue rather than corrupt conditionals.
         with engine.lock:
             engine._prepare(voice_path, exaggeration=req.exaggeration)
-            if req.format == "wav":
-                yield _wav_header(engine.sr)
-
             t_req = time.perf_counter()
             first_audio_ms: float | None = None
 
             if not req.stream:
+                # One-shot: the full length is known, so emit a correct-sized WAV
+                # header (a buffered consumer wants well-formed sizes).
                 wav, audio_sec = engine.generate(req.text, req)
                 first_audio_ms = (time.perf_counter() - t_req) * 1000
                 log.info("oneshot: %.0fms gen, %.2fs audio", first_audio_ms, audio_sec)
-                yield _pcm16(wav)
+                pcm = _pcm16(wav)
+                if req.format == "wav":
+                    yield _wav_header(engine.sr, len(pcm))
+                yield pcm
                 return
 
+            # Streaming: length is unknown up front → open-ended WAV header.
+            if req.format == "wav":
+                yield _wav_header(engine.sr)
+
             def _gen(text: str) -> tuple[torch.Tensor, float]:
                 return engine.generate(text, req)