ttlequals0/Audicle v0.56.7

0.56.7

on GitHub

0.56.7

Fixed

• A release that ships an unchanged base lexicon no longer re-imports it
  (#126 follow-up). The sync gate compared the stored value against the app
  version, so every version bump replayed all 1,336,791 rows even though the
  artifact was byte-identical; three releases in one day meant three full
  imports. The gate is now a sha256 of the artifact's bytes, which costs
  milliseconds to compute against the two minutes it saves. Upgrading
  re-imports once more to record the digest, then stays quiet until the
  lexicon data actually changes. The seed-resync migrations are unaffected:
  they call the importer directly and never consulted this gate.

0.56.6

Fixed

• The base-lexicon import no longer needs tens of gigabytes of transient
  disk (#126 follow-up). Turning off wal_autocheckpoint for the whole
  1.3M-row insert meant nothing could checkpoint until it finished, so the
  WAL grew without bound: measured here at 24 GiB peak for a 168 MB
  database. The insert now runs in 50,000-row batches with a truncating
  checkpoint between them, which caps peak WAL at 921 MiB with no change in
  import time (117s versus 112s). Since 0.56.4 was the first version to run
  this on the default startup path, a fresh install or version bump no
  longer demands disk headroom nothing announced.
• The import logs lexicon_sync_started and its completion with row counts
  and elapsed time, so its progress is visible from the logs instead of only
  by watching the WAL file grow.

0.56.5

Fixed

• Episode summaries and chapter titles can no longer ship AI-tell
  punctuation. The summary model kept emitting em dashes, en dashes, and
  curly quotes despite the prompt banning them, so the ban is now enforced
  deterministically after the LLM call: dashes become comma asides, numeric
  ranges keep a plain hyphen, quotes straighten, the ellipsis character
  expands, and emojis are dropped. Accented letters in names are untouched.
  The prompts also carry the ASCII rule in their top output contract now,
  not only in the style section.

0.56.4

Fixed

• The base-lexicon sync no longer holds migration_lock for its whole import
  (#126). On a fresh database or a version bump the import can run for tens
  of minutes on slow disks, and whichever process booted second sat blocked
  in its startup migration check the entire time, so the web process never
  bound port 8000. The sync now serializes on its own .lexicon-sync.lock;
  startup never waits on it. The import itself also got faster and lighter:
  a 64 MB page cache and no mid-transaction WAL checkpoints during the bulk
  insert (the 2 MB default cache wrote about 14 GB for a 241 MB artifact),
  rows inserted in key order for B-tree locality, and one deliberate
  checkpoint afterward. Reported with a full diagnosis in #126.
• Job-row writes (stage and progress updates) log a slow_db_write warning
  with the current WAL size when they exceed one second, to pinpoint the
  silent multi-minute stalls observed between pipeline stages on 0.56.3.

0.56.3

Fixed

• The normalize stage no longer spends minutes in the deterministic
  corrections backstop. lexicon.lookup()'s OR across an unindexed column
  full-scanned the base lexicon once per token (measured 100 to 144 ms per
  lookup in production, about 242 s per episode); the query is now two
  index-served probes with a new idx_lexicon_input_text index (migration
  026), 160x faster per lookup in benchmarks, and the per-token cache is
  shared across the whole episode instead of rebuilt per chunk. Instrumented
  runs put the pass at 92% of normalize; it should now be seconds.

0.56.2

Added

• The Recents list can now be pruned (#125). Each finished run's action menu
  gains "Remove from recents", and the RECENT header gains "clear failed" and
  "clear all" controls with an inline confirm step. All of it removes job rows
  only: queued and processing runs always survive, and episodes stay in the
  feed. New API: DELETE /api/v1/jobs/{id} and
  DELETE /api/v1/jobs?scope=all|failed.

0.56.1

Changed

• The default summary and chapter prompts now carry an anti-slop style
  section distilled from the humanizer rules: banned AI vocabulary and
  constructions (em dashes, colon reveals, forced triads, significance
  inflation, vague attribution, hedging stacks), a concrete-over-thematic
  rule, and an end-on-the-last-point rule; chapter titles additionally get
  sentence case and a name-the-topic rule. Deployments that edited either
  prompt in Settings keep their override; reset to defaults to pick up the
  new text.

0.56.0

Added

• A resumable TTS chunk cache: a synthesized chunk that clears the quality
  checks is stored under <data_dir>/tts_cache, keyed on backend, model,
  voice, language, text, and the baseline generation params, so a re-run of
  a job that failed partway through does not re-synthesize chunks it already
  produced. New settings TTS_CHUNK_CACHE_ENABLED (default on) and
  TTS_CACHE_RETENTION_DAYS (default 7). Caching is skipped whenever the
  active voice/model cannot be identified with confidence (no resolved
  voice slot, no configured model, an unseeded CHATTERBOX_SEED=0, or a
  failed model select). Each entry records whether it cleared the quality
  checks, so a run with audio analysis or ASR verification on ignores an
  entry stored by a run that had them off, and never inherits audio it
  would have checked.
• Adaptive max_chars: if 2 or more of a job's first 20 chunks needed a
  regeneration, the rest of the job synthesizes at a lowered
  CHATTERBOX_MAX_CHARS (75% of the configured value) instead of paying for
  a regen on most chunks. Lowering only; it is never raised back up mid-job. New
  setting TTS_ADAPTIVE_MAX_CHARS_ENABLED (default on). The lowered value is
  part of the TTS chunk cache key, so those chunks are still cached, under a
  key that describes what was actually synthesized rather than colliding with
  the unmodified-baseline entry for the same text.
• Idle wrapper restart: between jobs, when the queue is empty, the worker
  checks the tts-wrapper's /health (now reporting rss_mb and
  restart_recommended) and asks it to restart itself via the new
  POST /maintenance/restart if it's over its memory soft limit. That puts
  the restart on our schedule, instead of only ever restarting mid-job at the
  wrapper's hard limit. The wrapper rejects with 503 while inference is still
  running. New setting TTS_IDLE_RESTART_ENABLED (default on); a no-op for the
  openai-api TTS backend.
• A CI workflow. Pull requests and pushes to main now run ruff plus all three
  pytest suites (backend, render, tts-wrapper) and build
  the app image. Until now the full test and build gate was local-only; CI
  ran just CodeQL and dependency review.
• Each release now also publishes a CPU-only wrapper image,
  ttlequals0/audicle-tts:<version>-cpu, so a GPU-less deployment no longer
  has to build it locally. The trivy release gate scans the new tag with the
  tts ignorefile.
• An opt-in sentence-scoped pronunciation pass: with PRONUNCIATION_SCOPE
  set to sentence (default stays chunk), the pronunciation LLM call sends
  only the sentences that matched a reference term, as numbered lines, and
  splices the respelled sentences back in place. Any protocol or splice
  failure falls back to the existing full-chunk call, so the worst case is
  what happens today; a protocol failure also turns sentence scope off for the
  rest of that job, since a model that ignores the numbered-line contract once
  keeps ignoring it.
• A strict mode for remote ASR verification: WHISPER_API_STRICT (default
  off) fails a chunk instead of shipping it unverified when the openai-api
  ASR backend is configured but unreachable. The default keeps the existing
  degrade-to-unverified behavior.
• Per-chunk pipeline instrumentation in the logs: tts_chunk_done now
  carries synth_ms, attempts, rtf, and the wrapper-reported
  inference_ms, verify_ms, and rss_mb (also new in the wrapper's
  /generate response); pipeline_start reports queue_wait_seconds; the
  pronunciation pass logs pronunciation_window_done with llm_ms and
  input_chars. Real-time factor per chunk was previously not measurable
  from the logs at all.

Changed

• The README sample is a newer narration clip (#122). The download is
  docs/sample.mp4, the same file the player uses; docs/sample.mp3 is gone.
• TTS and remote-ASR calls now share one process-wide httpx.AsyncClient
  instead of opening a fresh one per chunk, and the remote SSRF guard caches a
  resolved host's verdict for 5 minutes instead of re-resolving DNS on every
  chunk.
• Episode audio is concatenated by streaming each chunk to disk with
  soundfile instead of building the whole episode as one in-memory tensor,
  removing the end-of-job memory spike on long episodes. Output is
  byte-identical to the old path.
• tts-wrapper/uv.lock now sources torch and torchaudio from the PyTorch CPU
  index, dropping the nvidia-*-cu12 wheel set (~7 GB) from any dev or CI sync
  of the chatterbox extra. The shipped images are unaffected: both wrapper
  Dockerfiles install torch with pip on their own.
• The CPU wrapper image no longer ships the build toolchain: build-essential
  and git are purged after the native extensions compile, which removes
  linux-libc-dev and with it 50 unfixable HIGH/CRITICAL kernel-header CVE
  findings from the trivy gate. The gate script itself now uses a 30 minute
  trivy timeout and reports a fatal scan error (cache lock, analysis timeout)
  as SCAN ERROR instead of conflating it with a CVE FAILED verdict.