From 5ca77f86b70718b4ef1a07299efbd6431268d546 Mon Sep 17 00:00:00 2001 From: historia Date: Thu, 20 Aug 2026 18:06:22 -0400 Subject: fix: do not chunk with audio.cpp backend (double chunking) --- converter/config.py | 12 ++++-------- 1 file changed, 4 insertions(+), 8 deletions(-) (limited to 'converter/config.py') diff --git a/converter/config.py b/converter/config.py index ed49f12..83778ab 100644 --- a/converter/config.py +++ b/converter/config.py @@ -7,14 +7,10 @@ API_TIMEOUT = 600 # Timeout per chunk request in seconds MAX_RETRIES = 3 # Attempts per chunk request HEARTBEAT_INTERVAL_SECONDS = 30 # Print "still working" in console logs every N seconds -# Words per TTS generation request. -# Note that qwen-tts-demo does no chunking at all, but faster-qwen3-tts and -# audio.cpp may do chunking as well, so you may be needlessly double-chunking. -# This is the only size limit: there is no hard ceiling. However, servers -# silently truncate audio when a single generation runs too long (roughly -# ~2.5 min on the faster backend's static KV cache, ~11 min on the Gradio -# demo) without reporting an error, so raising this is at your own risk. -# The client-side truncation check still catches and retries gross cases. +# Words per TTS generation request (client-side chunking). +# The gradio and faster backends always chunk with this size +# The audio.cpp backend chunks long text itself, so this is ignored +# by default with that backend. Force chunking with --chunk CHUNK_SIZE = 250 # Default TTS backend. -- cgit v1.2.3