Audiobook Producer
97 lines
4.2 KiB
YAML
97 lines
4.2 KiB
YAML
# One settings file for the whole pipeline. Relative paths are relative to the repo root.
|
|
|
|
books_dir: books # where book folders live: books/<first-author>/<book>/
|
|
work_dir: /tmp/audiobook-work # scratch only (OCR output, wav pieces). Point at local NVMe in production.
|
|
|
|
# Where finished audio goes. Empty = the book folder in the repo (the agreed default).
|
|
# Set to a folder to write <archive_dir>/<first-author>/<book>/audio there instead.
|
|
archive_dir: ""
|
|
|
|
narrator: "" # spoken in the credits only if filled in (leave empty for a TTS voice)
|
|
|
|
# Inserted parts. All OFF by default. Per-book and per-chapter switches in book.json override these.
|
|
summaries:
|
|
chapter_preview: false
|
|
chapter_summary: false
|
|
concept_summary: false # one mid-chapter concept summary
|
|
|
|
# OpenAI-compatible chat endpoint, used only when a summary switch is on.
|
|
llm:
|
|
endpoint: "" # e.g. http://host:8081/v1/chat/completions
|
|
model: ""
|
|
api_key_env: "" # name of an environment variable that holds the key (never put a key here)
|
|
timeout_s: 300
|
|
|
|
# OpenAI-compatible speech endpoint.
|
|
voice:
|
|
# LOCKED by Ty 2026-10-17 for MVP ("clip B"): Chatterbox, stock voice Michael, native settings only.
|
|
# Never set speed_factor: the server applies it as a time-stretch after generation and ruins the sound.
|
|
api: chatterbox_tts # chatterbox_tts = native /tts route; anything else = OpenAI-style /v1/audio/speech
|
|
endpoint: http://voice:8004/tts # override with env AUDIOBOOK_VOICE_ENDPOINT (CI uses the LAN address)
|
|
model: chatterbox
|
|
voice: Michael.wav
|
|
extra: # extra JSON fields sent with every request
|
|
exaggeration: 0.25
|
|
cfg_weight: 0.2
|
|
temperature: 0.6
|
|
seed: 7
|
|
split_text: true
|
|
chunk_size: 240 # the server splits long text at sentence ends into chunks of about this size
|
|
timeout_s: 600
|
|
retries: 3 # a failed request is retried this many times (5 s, then 15 s, then 45 s apart) before the run fails
|
|
load_if_needed: true # if the server is up but its model is unloaded, ask it to load (POST /restart_server) before rendering
|
|
unload_when_done: false # true = free the GPU after a render (POST /api/unload); the next run loads it again
|
|
max_chunk_chars: 100000 # one request per paragraph; the server does the sentence splitting (chunk_size above)
|
|
gap_sentence_s: 0.25 # pause between pieces inside a paragraph
|
|
gap_paragraph_s: 0.7 # pause between paragraphs and after cues and labels
|
|
|
|
ocr:
|
|
enabled: true # run ocrmypdf only when a page has no text layer
|
|
required: true # pages without a text layer and no working OCR = the run fails (env AUDIOBOOK_REQUIRE_OCR=0 to allow)
|
|
language: eng
|
|
min_chars_per_page: 10 # fewer characters than this = no text layer
|
|
|
|
meta:
|
|
first_pages: 10 # title/author are looked for on the first pages (OCR only for pages with no text layer)
|
|
|
|
fetch: # a .url/.txt file in _incoming/ holding a link to a PDF
|
|
max_mb: 300 # larger downloads are refused
|
|
connect_timeout_s: 30
|
|
max_time_s: 900 # a whole download may take this long
|
|
|
|
script:
|
|
retail_sample_words: 400 # about 2.5 minutes at 155 words per minute (ACX asks for 1 to 5 minutes)
|
|
chime_label: Chime
|
|
|
|
license:
|
|
public_domain_max_year: 1930 # US, as of 2026
|
|
|
|
audio:
|
|
sample_rate: 44100
|
|
channels: 1
|
|
mp3_bitrate: 192k
|
|
head_room_tone_s: 0.75 # ACX 0.5 to 1
|
|
tail_room_tone_s: 2.0 # ACX 1 to 5
|
|
room_tone_db: -66 # RMS level of the room tone (ACX noise floor limit is -60)
|
|
target_rms_db: -20.5 # ACX -23 to -18
|
|
peak_ceiling_db: -4.0 # limiter ceiling; ACX limit is -3 and MP3 encoding can add a little
|
|
chime_hz: 660
|
|
chime_s: 1.0
|
|
chime_db: -18 # peak level of the chime in the final file, kept soft
|
|
|
|
acx: # limits used by the checker (R13)
|
|
min_bitrate_kbps: 192
|
|
sample_rate: 44100
|
|
rms_min_db: -23
|
|
rms_max_db: -18
|
|
peak_max_db: -3
|
|
noise_floor_max_db: -60
|
|
head_min_s: 0.5
|
|
head_max_s: 1.0
|
|
tail_min_s: 1.0
|
|
tail_max_s: 5.0
|
|
max_minutes: 120
|
|
retail_min_s: 60
|
|
retail_max_s: 300
|
|
speech_window_db: -50 # a 50 ms window louder than this counts as speech when measuring head and tail
|