feat: reuse one seed per run for consistent voice across chunks

1 files changed, 16 insertions, 28 deletions

diff --git a/converter/config.py b/converter/config.py
index a19ac63..9f4dd95 100644
--- a/converter/config.py
+++ b/converter/config.py
@@ -10,53 +10,41 @@ file formats) is fixed in the code where it is used.
 # /run_instruct); the faster backend is the OpenAI-compatible server from
 # the faster-qwen3-tts repository.
 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
+CLONE_API_URL = "http://127.0.0.1:7861"  # Base-model demo
+FASTER_API_URL = "http://127.0.0.1:8000"  # faster-qwen3-tts server
 
 # 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)
+CHUNK_SIZE = 250
 
 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
 
 LANGUAGE = "English"
+
+# Seed sent to the TTS servers (only the endpoints that accept one; the
+# primary /run_instruct and /run_voice_clone endpoints and the faster
+# backend never receive a seed). -1 means "randomize per generation".
+# With CONSTANT_SEED = True and SEED = -1, one random seed is drawn at
+# startup and reused for every request of the run, keeping the voice
+# consistent across chunk boundaries; set SEED to a fixed number to also
+# reproduce the same voice across runs.
 SEED = -1
+CONSTANT_SEED = True
 
-CUSTOM_VOICE_SPEAKER = "Vivian"
-CUSTOM_VOICE_INSTRUCT = "Speak naturally and clearly, as if reading a dramatic book to an adult audience."
+SPEAKER = "Vivian"
+INSTRUCT = "Speak naturally and clearly, as if reading a dramatic book to an adult audience."
 
-VOICE_CLONE_USE_XVECTOR_ONLY = False
+XVECTOR_ONLY = False
 
 # Must match a key in the faster server's voices.json ("default" when the
 # server was launched with --ref-audio). The server silently falls back to
 # its first voice for unknown names.
-FASTER_TTS_VOICE = "default"
+FASTER_VOICE = "default"
 
 AUDIO_FORMAT = "m4b"  # Default output container
 AUDIO_BITRATE = "128k"