Skip to content

CHANGELOG.md

Latest commit

 

History

History
110 lines (92 loc) · 8.45 KB

CHANGELOG.md

File metadata and controls

110 lines (92 loc) · 8.45 KB
Raw

Changelog

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

This project follows semantic versioning once published to PyPI.

Unreleased

Fixed

  • Preserved exact digital silence in lavasr-compat instead of passing it through the model and producing deterministic low-level output texture.
  • Made audio-super-res eval matrix return a non-zero exit status when any matrix run fails, so remote evidence workflows cannot report success for failed stability checks.
  • Corrected load_time_seconds and total_elapsed_seconds comparison direction to lower_is_better alongside the other runtime metrics.

Added

  • Added a versioned release regression threshold policy and a Colab CLI workflow for same-backend baseline/candidate matrix evidence.
  • Added a pinned, license-recorded LibriSpeech dev-clean tiny baseline specification and Colab-only real-speech matrix workflow.
  • Added a pinned external Whisper evaluator specification and Colab-only transcript workflow for precomputed downstream WER/CER evidence without adding an ASR package dependency.
  • Added a Colab-only verified MUSHRA bundle export for the licensed LibriSpeech sinc/LavaSR matrix.
  • Added a bounded Colab-only MossFormer2 feasibility workflow covering minimal checkpoints, safe conversion, offline CPU inference, channel behavior, alignment, and parity-fixture constraints.
  • Added a Colab T4 LavaSR eager/torch.compile benchmark and recorded the evidence-based decision to keep eager as the only selectable provider.

0.6.0 - 2026-07-12

Added

  • Added an experimental self-contained LavaSR-compatible torch runtime with Vocos-style modules, strict checkpoint loading, and verified local weight resolution.
  • Added offline LavaSR v2 bundle validation, environment-gated real-weight download verification, torch smoke coverage, and upstream LavaSR/Vocos parity testing.
  • Added golden-sample comparison helpers, tests, and fixture documentation for compatible backends.
  • Expanded model catalog metadata and added a public model admission scorecard API.
  • Added v0.3.0 model-planning scope and v0.4.0 accelerator/runtime-provider planning scope.
  • Added speech and general-audio candidate reviews for ClearerVoice, Resemble Enhance, AudioSR, FlowHigh, and deferred research candidates.
  • Added a repository-based LavaSR/GPU Colab validation guide and evidence checklist.
  • Added a fresh Colab T4 validation record for lavasr-compat real-weight CUDA inference.
  • Added format-check, typecheck, package metadata, and wheel contents checks to the Pixi quality tasks and CI/release workflows.
  • Added a PEP 561 py.typed marker for downstream type checkers.
  • Completed v0.3.0 feasibility conclusions for ClearerVoice MossFormer2_SR_48K, FlowHigh, and Resemble Enhance, deferring each until its validation, checkpoint, or runtime blockers are resolved.
  • Added accelerator/runtime-provider metadata, provider resolution helpers, and benchmark JSON output for gated hardware validation.
  • Added accelerator installation, validation-matrix, and LavaSR optimization guidance.
  • Added the first lightweight backend evaluation harness with controlled degraders, full-reference metrics, eval manifests, and audio-super-res eval run/compare.
  • Added engineering/stability fields to eval manifests and benchmark JSON, including failure status, failure-case classification, RTF/peak RSS reporting, and backend capability/governance profiles.
  • Added threshold-based audio-super-res eval compare checks and comparison tables for audio quality, downstream, engineering, stability, and governance evidence.
  • Documented optional PESQ/STOI/ESTOI/MCD and codec-degrader integration requirements for full-reference evaluation.
  • Added audio-super-res eval no-reference with builtin CPU/offline signal-stat screening records and documented gated DNSMOS/NISQA/UTMOS/ViSQOL adapter requirements.
  • Added audio-super-res eval downstream with builtin ASR transcript WER/CER delta evaluation and gated speaker/VAD/KWS adapter schema.
  • Added audio-super-res eval listening-export for AB/ABX/MUSHRA-ready blind bundles with separate answer keys and explicit rating dimensions.
  • Completed the v0.5.0 evaluation/regression harness milestone scope across full-reference, no-reference, downstream, listening, engineering/stability, and regression workflows.
  • Prepared source version metadata for 0.6.0 release hardening after the completed v0.2-v0.5 milestones.
  • Added audio-super-res eval init-speech-bwe to generate a deterministic synthetic speech_bwe_v1_tiny evalset for smoke/regression workflows.
  • Added deterministic opus_16k_24kbps and mp3_32kbps codec-like degraders for lightweight eval coverage without requiring ffmpeg.
  • Added explicit optional full-reference adapter attempts through eval run --optional-metric for PESQ, STOI, ESTOI, and MCD, with skipped dependency records when unavailable.
  • Added backend load/init time and total elapsed time fields to benchmark and eval performance reports.
  • Added audio-super-res eval matrix to run backend/degrader grids and write a matrix.json index over individual eval manifests.
  • Added matrix comparison, JSON threshold policies, Markdown eval reports, artifact bundles, dataset manifest validation, lightweight MCD, optional full-reference extras, and optional CUDA memory profiling metadata.
  • Added matrix --reuse-existing / --fail-fast, synthetic failure-case evalset generation, explicit local torch checkpoint inspection, and external backend adapter protocol documentation.

Changed

  • Wired lavasr-compat past verified local weights into the experimental runtime while keeping provider downloads outside backend inference.
  • Removed PyYAML from the lavasr extra; the extra now only provides the torch runtime dependency.
  • Reworked lightweight golden/preprocessing spectral paths to avoid fragile SciPy stateful filtering/STFT calls in default tests.
  • Updated roadmap and release documentation for the completed v0.1.x release flow and completed v0.2.0 validation gates.
  • Updated model and backend listing metadata to report declared accelerators and runtime providers without importing heavy runtimes.
  • Simplified documentation ownership by adding a docs index and keeping README focused on user entry points.
  • Consolidated GPU/Colab, golden validation, backend candidate, and release records into their long-lived owner documents; refocused the roadmap on post-v0.6 priorities.
  • Updated Docker installation to use uv pip install --system instead of direct pip install.

Fixed

  • Fixed eval matrix run-manifest resolution when matrix.json was created from a relative output directory and later compared or bundled from a different working directory.

0.1.1

  • Fixed Python 3.10 import compatibility by replacing Python 3.11-only datetime.UTC usage.
  • Superseded the yanked 0.1.0 PyPI files as the current public alpha baseline.

0.1.0

  • Added the initial Pixi-managed Python package.
  • Added audio-super-res and audiosr CLI entry points.
  • Added single-file and recursive directory batch processing.
  • Added explicit chunked processing with overlap crossfading for long files.
  • Added backend selection with the baseline sinc-resample backend.
  • Added inference configuration for device, precision, chunking, seed, and model cache path.
  • Added audio quality checks for sample rate, duration drift, clipping, and peak level.
  • Added standalone JSON quality report export through --quality-report-json.
  • Added examples for single-file enhancement, batch processing, and quality checks.
  • Added the optional audiosr backend wrapper for AudioSR latent diffusion inference.
  • Added an environment-gated real AudioSR integration test.
  • Added backend availability metadata and JSON output for --list-backends.
  • Added model catalog output through --list-models.
  • Added JSON manifests for dry-run plans and completed enhancement runs.
  • Added manifest regression comparison through --compare-manifests.
  • Added optional low-pass preprocessing for model-backed enhancement runs.
  • Added managed weight manifests, local cache verification, and explicit Hugging Face download plumbing.
  • Added LavaSR-compatible managed weight metadata while keeping self-contained LavaSR inference marked as pending.
  • Added PyPI Trusted Publishing / GitHub OIDC release workflow configuration.
  • Added release dry-run notes, a Colab plan, and sample JSON artifacts.
  • Added GitHub issue templates, Dependabot configuration, CodeQL security workflow, and architecture/test documentation.