fix: limit chunk size to 250

1 files changed, 25 insertions, 4 deletions

diff --git a/converter/config.py b/converter/config.py
index eb01462..a19ac63 100644
--- a/converter/config.py
+++ b/converter/config.py
@@ -13,14 +13,35 @@ QWEN_API_URL = "http://127.0.0.1:7860"  # CustomVoice demo
 VOICE_CLONE_API_URL = "http://127.0.0.1:7861"  # Base-model demo
 FASTER_TTS_API_URL = "http://127.0.0.1:8000"  # faster-qwen3-tts server
 
-# Words per TTS generation request. Each API call is ONE model generation:
-# long generations lose prosody and can degrade into garbled audio.
-CHUNK_SIZE_WORDS = 40
+# Words per TTS generation request. Each request is ONE model generation:
+# the voice is re-sampled per request (every chunk boundary can drift
+# slightly), while over-long generations lose prosody and can turn garbled.
+# ~250 words is ~1.5-2 minutes of speech: few voice boundaries while
+# staying inside both servers' generation caps.
+CHUNK_SIZE_WORDS = 250
+
+# Hard ceiling on words per request, regardless of CHUNK_SIZE_WORDS. Both
+# TTS servers silently truncate audio when a single generation exceeds its
+# cap (~2.5 min for the faster backend's static KV cache, ~11 min for the
+# Gradio demo) without reporting any error, so larger requests are always
+# split client-side. Keep a margin below ~300 words to survive slow
+# narration on the faster backend.
+MAX_REQUEST_WORDS = 250
+
+# Duration sanity check: a response whose audio is far shorter than its
+# word count implies is treated as silently truncated, fails the request,
+# and goes through the normal retry logic. 150 wpm is a typical spoken
+# pace; the ratio is set low (0.5) so only gross truncation trips it.
+ESTIMATED_WORDS_PER_MINUTE = 150
+MIN_AUDIO_DURATION_RATIO = 0.5
+MIN_WORDS_FOR_DURATION_CHECK = 10
+
 VOICE_CLONE_MAX_CHUNK_CHARS = 200  # Server-side re-chunking limit for clone requests
 VOICE_CLONE_CHUNK_GAP = 0  # Pause (seconds) between server-side clone chunks
+
 MIN_DELAY_BETWEEN_CHUNKS = 0  # Pause between API calls (rate-limit protection; local servers need none)
 
-API_TIMEOUT = 300  # Seconds before an API call times out
+API_TIMEOUT = 600  # Seconds before an API call times out (a ~250-word request can take minutes on the Gradio demo)
 MAX_RETRIES = 3  # Attempts per chunk request
 HEARTBEAT_INTERVAL_SECONDS = 30  # Print "still working" this often during a chunk