chat-with-audio CLAUDE.md

Chat with Audio — development notes

Chat-driven audio enhancement tool: MCP server (FastMCP, stdio) + C++ DSP core (pybind11) + local A/B viewer. See README.md and docs/ for user documentation (tool reference, workflows, compliance, smart regions, recipes, architecture).

Name everywhere: Chat with Audio (package chat_with_audio, MCP server chat-with-audio, GitHub chat-with-audio, local project folder "Chat with Audio").

Commands

uv sync --all-extras                              # build (incl. C++) + all deps
uv sync --reinstall-package chat-with-audio # after changes in cpp/
uv run pytest                                     # test suite
uv run ruff check .                               # lint (also runs in CI)
uv run python scripts/mcp_smoke.py                # MCP stdio smoke test
uv run ait analyze <file>                         # dev CLI without MCP
uv run ait improve <file> [--profile speech|music] [--denoise-method ai]
uv run ait viewer                                 # viewer on :8471

Architecture map

• cpp/ — header-only DSP (biquad.hpp, dynamics.hpp) + bindings.cpp → module chat_with_audio._dsp. Arrays are float32 (channels, n); functions return new arrays; dynamics use a linked detector across channels.
• src/chat_with_audio/dsp/__init__.py — dispatch: native _dsp when built, otherwise fallback.py (scipy; identical signatures, block-based dynamics). spectral_nr.py = Tier A denoise; ai_nr.py = Tier B (DeepFilterNet).
• analysis.py → metrics dict + score_and_issues(); improve.py → profile detection + rules → (steps, rationale); chain.py → STEP_REGISTRY + execution (incl. leveler and segment-driven smart_denoise).
• regions.py → smart problem regions: windowed detectors (hum/noise/clip/boom) find where on the timeline something is wrong; per-region mini-chains are applied with raised-cosine crossfades, everything outside stays untouched (smart_edit tool). Noise reference floor is clamped to -80 dB; boom regions inside a hum region are dropped (the notch already covers them).
• recipes.py → saved chains as shareable JSON (save_recipe/apply_recipe/ list_recipes); built-ins live in src/chat_with_audio/recipes/, user recipes in ~/AudioImprove/recipes/ (env AIT_RECIPES_DIR; tests isolate this automatically). chain.validate_steps() guards every load/save.
• segments.py → speech/music/silence segmentation (level-Otsu primary; modulation rhythm as fallback). refine.py → iterative measure-and-adjust loop (refine_audio tool): AI denoising once up front, then adjust leveler/loudness until the speech peak and balance are right; silence segments are pushed back down afterwards (_duck_silence) because the leveler would otherwise lift them.
• compliance.py → delivery-spec registry (EBU R128, ATSC A/85, Netflix 2.0 én 5.1 dialogue-gated, streaming, ACX) + pass/fail checker incl. formaat- en kanaaleisen; master_for schrijft compliance.json (viewer-paneel) en exporteert mono → dual-mono bij een 2.0-eis. dialogue_loudness = blok- gebaseerde spraak-gated meting (DI-achtig, niét het Dolby-algoritme); op 5.1 detectie op het centerkanaal.
• Surround: SURROUND_LAYOUTS in analysis.py — 5.1 (SMPTE) krijgt gewogen BS.1770 (LFE eruit via loudness_view), per-kanaal-QC en ITU-downmix-piek; ADM BWF (Atmos-metadata) wordt herkend via axml/chna-chunks in io.probe.
• dsp/dialogue.py → breath_control / deplosive / duck_music (chain steps); gain envelopes are smoothed with edge padding — plain convolution would drag file edges toward zero.
• markers.py → region map → Audition marker CSV + Audacity labels + JSON (export_markers).
• dsp/roomtone.py → room-tone fill: digitale gaten vullen met geshuffelde overlap-add van de eigen ambience (fill_room_tone); qcsheet.py → markdown-QC-rapport (qc_report, batch: qc_folder).
• dsp/spectral_repair.py → spectral painting (spectral_repair): magnitudes interpoleren uit de context, fase phase-vocoder-coherent voortzetten (bin-centerfrequenties laten mainlobe-bins driften — gemeten dphi gebruiken).
• duck_music heeft twee modi: beds (segmentniveau, licht) en stems (Demucs-sidechain voor muziek ónder spraak, [stems]-extra).
• sync.py → 32-sporenrecorder (sync_tracks): envelope-GCC-PHAT + full-rate verfijning, confidence per spoor, klokdrift-meting/-correctie; uitgelijnde wavs + .sesx; A/B-sessie = ongesynct vs gesynct mixdown. Valkuil: strak periodiek materiaal (metronoom) is inherent dubbelzinnig voor correlatie — testsignalen moeten aperiodiek gaten (en recorder-seeds ver van event-seeds, anders ontstaat een echte schijncorrelatie).
• server.py — 30 MCP tools; sessions.py — session folders under ~/AudioImprove/sessions/ (env AIT_SESSIONS_DIR; tests isolate this automatically). Every session writes timeline.json (segments + treated regions) for the viewer's timeline lane; ids get a -2 suffix on collision.
• viewer/server.py — stdlib http.server on 127.0.0.1:8471 (env AIT_VIEWER_PORT); viewer/static/app.js — A/B player: both buffers always play together, switching = gain crossfade.