feat: audio.cpp backend support

1 files changed, 47 insertions, 39 deletions

diff --git a/converter/config.py b/converter/config.py
index 9f4dd95..4cb5ab8 100644
--- a/converter/config.py
+++ b/converter/config.py
@@ -1,50 +1,58 @@
-"""Configuration for the audiobook converter.
-
-Edit these values to change voices and processing behavior. Everything
-else (voice mode names, languages, speaker names, model ids, folders,
-file formats) is fixed in the code where it is used.
-"""
-
-# Server endpoints. Voice clone needs the Base-model demo, which is a
-# separate server from the CustomVoice demo (that one only exposes
-# /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
-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.
+# Default output options
+AUDIO_FORMAT = "m4b"
+AUDIO_BITRATE = "128k"
+LANGUAGE = "English"
+
+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-qwen-tts and
+# audio.cpp may do chunking as well, so you may be needlessly double-chunking.
 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
+# Default TTS backend.
+# gradio: qwen-tts-demo
+# faster: faster-qwen-tts
+# audiocpp: audiocpp_server
+# The --backend CLI flag overrides this
+BACKEND = "gradio"
 
-LANGUAGE = "English"
+###############################################################################
+# BACKEND 1: qwen-tts-demo (gradio) options                                   #
+###############################################################################
 
-# 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
+# There are different API URLs for CustomVoice and Base models so you can run both at once
+QWEN_API_URL     = "http://127.0.0.1:7860"  # CustomVoice model
+CLONE_API_URL    = "http://127.0.0.1:7861"  # Base model
 
-SPEAKER = "Vivian"
+# Custom voice options
+SPEAKER = "Vivian"  #Vivian, Serena, Uncle_Fu, Dylan, Eric, Ryan, Aiden, Ono_Anna, Sohee
 INSTRUCT = "Speak naturally and clearly, as if reading a dramatic book to an adult audience."
 
+# Don't clone with transcription, only use x-vector-only cloning. Generally "worse"
 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.
+# Randomization seed. -1 means randomize with every generation
+# With SEED = -1 and CONSTANT_SEED = True, one random seed will be used for the entire audiobook.
+# This may keep the voice slightly more consistent across chunk boundaries
+SEED = -1
+CONSTANT_SEED = False
+
+###############################################################################
+# BACKEND 2: faster-qwen-tts options                                          #
+###############################################################################
+FASTER_API_URL   = "http://127.0.0.1:8000"    # faster-qwen3-tts server (Base model only)
+
+# Default voice if no --voice is passed
 FASTER_VOICE = "default"
 
-AUDIO_FORMAT = "m4b"  # Default output container
-AUDIO_BITRATE = "128k"
+###############################################################################
+# BACKEND 3: audio.cpp options                                                #
+###############################################################################
+AUDIOCPP_API_URL = "http://127.0.0.1:8080"  # audio.cpp audiocpp_server
+
+# Model ids in the audio.cpp server.json config.
+AUDIOCPP_MODEL_ID = "qwen"https://github.com/0xShug0/audio.cpp
+AUDIOCPP_CLONE_MODEL_ID = "qwen-clone"