Skip to content

CHANGELOG.md

Latest commit

 

History

History
127 lines (87 loc) · 10.4 KB

CHANGELOG.md

File metadata and controls

127 lines (87 loc) · 10.4 KB
Raw

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[0.2.2] - 2026-05-22

Changed

  • Word-level speaker in combined transcripts now falls back to the enclosing segment's speaker when WhisperX's word-level diarization join doesn't intersect any pyannote span (typically a boundary-overlap gap on short, low-confidence trailing words). Previously such words were emitted with speaker: null despite the segment carrying a confident label. The Pydantic contract is unchanged — word-level speaker is still nullable in principle — but in practice the field is now populated whenever the segment-level assignment is confident.

Added

  • Published JSON Schemas for the combined transcript and batch summary documents under docs/schemas/ (combined-transcript-v1.json, batch-summary-v1.json). They are generated from the Pydantic source-of-truth in serialization mode, attached as GitHub release assets, and kept in lockstep with the models by a CI drift check. Consumers without access to the Python package can now fetch the schema for a pinned AR version directly from the release tag and validate transcripts against it. See docs/service.md for the consumption pattern.
  • make generate-schemas and make check-schemas Makefile targets. The drift check now runs as part of make all-checks and as a dedicated CI job (Schema Drift Check in ci.yml).

Documentation

  • Refreshed the README banner image.

[0.2.1] - 2026-05-21

Fixed

  • Service mode now defaults CUDA_DEVICE_ORDER to PCI_BUS_ID at startup (matching the CLI, respecting an explicit override), so REFINERY_DEVICE=cuda:N selects the GPU at nvidia-smi index N rather than CUDA's default FASTEST_FIRST ordering. Previously a multi-GPU host could load models onto the wrong card.
  • WhisperX forced-alignment failures are now logged at debug level (with the detected language and device) before falling back to transcription-only segments, instead of being swallowed silently.

Documentation

  • Corrected the service-mode tmpfs example to --mount type=tmpfs,dst=/scratch,tmpfs-mode=1777. A bare --tmpfs /scratch mounts root-owned, so the non-root refinery user could not write scratch and every job failed with PermissionError. Applied the same fix in docs/deployment.md.
  • Added service-mode GPU-selection guidance, an end-to-end local-directory batch runbook, and troubleshooting entries (tmpfs permissions, wrong-GPU selection, batch-size cap) to docs/service.md.
  • Added the project banner to the README.

Security

  • Upgraded vulnerable transitive and dev dependencies to patched versions: aiohttp 3.13.3 → 3.13.5, requests 2.32.5 → 2.34.2, Pygments 2.19.2 → 2.20.0, and pytest 9.0.2 → 9.0.3. Closes 13 Dependabot alerts (10 aiohttp plus requests, Pygments, pytest). The torch and transformers alerts remain deferred to v0.3.0, when WhisperX removal lifts the torch==2.1.2 / transformers<4.40 pins.

[0.2.0] - 2026-05-20

Added

  • HTTP service mode: a long-lived FastAPI service (audio-refinery-service entry point) exposing POST /transcribe, GET /jobs/{id}, and GET /health. Accepts multi-job batches, processes them serially on a background worker, and writes one combined transcript per job plus one summary per batch. Models load once at container startup and stay resident across jobs.
  • Combined transcript schema (v1.0.0) and per-batch summary schema (v1.0.0) — service-mode output documents that envelope the existing per-stage results so consumers read a single document per job/batch.
  • Bearer-token authentication via the REFINERY_API_KEYS allowlist; request and job-lifecycle logs carry a non-reversible caller fingerprint.
  • URI-driven I/O for the service: https:// (presigned GET for input, presigned PUT for output/summary) and file:// for local-dev and shared-volume deployments.
  • Multi-stage Dockerfile (CUDA 12.1 runtime base, non-root refinery user) published as lunarcommand/audio-refinery on Docker Hub.
  • audio-refinery serve CLI subcommand — a convenience wrapper around the service entry point.
  • Optional GPU thermal guard in service mode via REFINERY_GPU_TEMP_LIMIT / REFINERY_GPU_TEMP_POLL_SECONDS, reusing the CLI's monitor.
  • Service configuration via environment variables: REFINERY_API_KEYS, REFINERY_DEVICE, REFINERY_WHISPER_MODEL, REFINERY_COMPUTE_TYPE, REFINERY_DEFAULT_LANGUAGE, REFINERY_SENTIMENT_ENABLED, REFINERY_SCRATCH_DIR, REFINERY_INTERMEDIATE_DIR, REFINERY_MAX_BATCH_SIZE, REFINERY_MAX_QUEUE_SIZE, REFINERY_JOB_RETENTION_SECONDS, REFINERY_PORT, and REFINERY_LOG_FORMAT.

Changed

  • Demucs scratch directory is now host-agnostic: it defaults to tempfile.gettempdir()/audio-refinery-demucs and honors REFINERY_SCRATCH_DIR in both CLI and service mode, replacing the hard-coded /mnt/fast_scratch path.
  • CLI batch input now accepts any *.wav file; the audio_ filename prefix is optional and stripped from the derived content ID, replacing the previous required audio_<id>.wav pattern.
  • Sentiment analysis treats segments with no transcribed speech (e.g. silent audio) as no-ops rather than failures.
  • Added fastapi, uvicorn[standard], and httpx as main dependencies for service mode.
  • Documentation restructured into a two-path (CLI + service) layout: lowercase docs/ filenames, a slimmed README with a "choose your path" landing, and new docs/cli.md, docs/service.md, and docs/index.md.

Security

  • Scoped both GitHub Actions workflows (ci.yml, release.yml) to least-privilege permissions. Workflow default is now contents: read; the github-release job retains an explicit contents: write override. Closes CodeQL workflow-permission alerts.
  • Upgraded vulnerable transitive dependencies: idna 3.11 → 3.15, Mako 1.3.10 → 1.3.12, Pillow 12.1.1 → 12.2.0, urllib3 2.6.3 → 2.7.0. Closes nine Dependabot alerts covering path-traversal, redirect-header forwarding, decompression-bomb, OOB-write, and integer-overflow CVEs.

Known security debt

  • 18 Dependabot alerts remain open against torch and transformers (including one critical and several high severity). Both packages are pinned to WhisperX-compatible versions (torch==2.1.2, transformers>=4.30,<4.40) and cannot be upgraded without breaking the current ASR pipeline. These alerts close together when v0.3.0 lands and removes WhisperX from the critical path; see the universal alignment plan for the migration path.

Tests

  • Suppressed live Slack webhook calls during the test run via an autouse _no_slack fixture in tests/conftest.py. Previously, notifier._send() loaded the developer's .env on every invocation and POSTed real messages whenever pipeline tests triggered end-of-batch notifications.

0.1.1 - 2026-03-06

Added

  • combined_report.json now includes four derived metrics: avg_time_per_file_seconds, avg_time_per_mb_seconds, processing_speed_ratio (real-time factor), and words_per_audio_hour (transcription density)
  • Slack notifications now include detailed per-stage stats (processed / skipped / failed counts) and average processing time per file
  • make test-slack Makefile target for validating Slack webhook integration
  • Dockerfile and .dockerignore for containerized deployment
  • Sentiment output directory (<base>/sentiment/) support in batch pipeline

Changed

  • Centralized Demucs scratch directory resolution in CLI — RAM disk detection and fallback confirmation now happen in one place
  • Worker status reporting and failure aggregation in pipeline-parallel refactored for improved accuracy
  • python-dotenv import in Slack notifier is now conditional — avoids import-time failure when the package is absent
  • DEPLOYMENT.md expanded: Hugging Face token setup, NVIDIA driver requirements, cloud instance guidelines, and Docker usage
  • Combined report fields documented in README under the Parallel Pipeline section

Fixed

  • Narrowed exception handling in gpu_utils.py, transcriber.py, and notifier.py to avoid masking unexpected errors
  • Typo in SeparationError docstring

0.1.0 - 2026-03-01

Added

  • Initial release
  • separate command: GPU-accelerated vocal separation via Demucs htdemucs
  • diarize command: Speaker diarization via Pyannote speaker-diarization-3.1
  • transcribe command: Transcription with word-level alignment via WhisperX large-v3
  • sentiment command: Per-segment sentiment analysis via cardiffnlp/twitter-roberta-base-sentiment-latest
  • pipeline command: Single-GPU batch processing through all four stages
  • pipeline-parallel command: Multi-GPU batch processing with worker distribution
  • GPU pre-flight checks via nvidia-smi with active process detection
  • Thermal monitoring with configurable shutdown threshold (default 80°C)
  • Slack webhook notifications for pipeline completion and thermal events
  • Resume behavior for batch pipelines (skips already-completed files)
  • Pydantic data models for all pipeline outputs with full provenance tracking
  • Rich terminal output with progress spinners and result tables
  • VRAM usage tracking and peak memory reporting per stage
  • Scratch space management with automatic cleanup of intermediate files
  • GPU performance-based ordering via gpu_tflops.toml lookup table

Fixed

  • WhisperX model loading: pinned install to commit 741ab9a (v3.7.6) — the v3.1.1 tag uses an older API (transcribe.py) that lacks the device_index parameter required by the ctranslate2 backend
  • make install-whisperx now also installs WhisperX runtime deps (av, ctranslate2, faster-whisper, flatbuffers, nltk, onnxruntime) which were previously missing after a --no-deps install
  • transformers capped at <4.40.0 — versions 4.40+ use torch.utils._pytree.register_pytree_node, an API introduced in PyTorch 2.2, which breaks with the pinned PyTorch 2.1.2
  • make dev-setup now reinstalls CUDA torch wheels (torch==2.1.2+cu121, torchaudio==2.1.2+cu121) as its final step — uv sync resolves torch from PyPI and installs the CPU-only build, silently breaking GPU inference