Audio MCP

Measure an audio file (or the start..end section): peak/true-peak/RMS/LUFS loudness, clipping, DC offset, a loudness sparkline over time, silence and noise floor, stereo correlation/width/balance, spectral balance (band energy, brightness, tonal vs noisy, strongest frequencies as notes), pitch, onsets/tempo, start/end clicks and loop-seam quality. Ends with `issues`: plain-language problems ranked by severity. Set pitch/rhythm to false to run faster.

Apply one operations chain (same format as audio_process) to every file in a folder or list, writing results with the same relative names into output_dir. Refuses to write into the input folder unless overwrite=true. Example: normalize a set of sound effects to the same loudness: [{"op":"trim_silence"}, {"op":"normalize_lufs","lufs":-18,"limit":true}]. subtype: sample format for every output, e.g. PCM_16, PCM_24, FLOAT, or PCM_U8 for 8-bit WAV (retro or low-memory targets: combine with to_mono and resample). Checked before any file is written.

Analyze a folder (recursive; filter with a glob `pattern` like "*.wav" or "footstep*") or a list of files, and report how consistent they are: loudness outliers, mixed sample rates/channels/formats, inconsistent leading silence, files with problems. Includes `suggested_fixes` with ready-to-use operations for audio_batch_process, and a per-file table.

Check a file against a delivery spec and get a verdict (pass / pass with warnings / fail), one entry per criterion (required, measured, status, note), and `fix`: an ordered audio_process operations list that was simulated on this file (`fix.verified`), plus `cannot_fix` for problems no processing can solve (clipping, a noise floor too high to remove without damaging speech, wrong length). Specs: apple_podcasts, spotify, youtube, ebu_r128, atsc_a85, acx_audiobook, game_sfx (spec="list" describes them). Each result cites its sources and the date they were checked. After a fix that includes denoise, verify the words with audio_transcribe.

Audit the audio of a Godot 4 project (the folder with project.godot): inventory of every .wav/.ogg/.mp3 with a category (ui, sfx, voice, music, ambience, inferred from folder/file names; override with `categories`, e.g. {"audio/stingers": "music", "*/click_*": "ui"}), import settings from the .import files (music/ambience that doesn't loop, SFX/UI that do, clicky loop seams, never-imported files), loudness consistency per category (one-shots by LUFS momentary max, music/ambience by integrated LUFS; outliers, median, spread), per-file problems (clipping, leading silence, clicky ends, DC offset, mixed sample rates, stereo SFX) and res:// references in .tscn/.tres/.gd/.cs files (broken paths, wrong letter case, unreferenced files). Returns `summary`, per-category `categories` stats with `suggested_batch` for audio_batch_process, and `issues` sorted problem > warning > info, each with a res:// file and a `fix` (audio_process operations or an editor step). res://x means <project_dir>/x. Read-only: never modifies the project. Analyzes at most max_files files, each over its first max_seconds (loop seams use the real loop points).

A/B two files: plain-language summary, metric deltas (loudness, peak, brightness, tempo, pitch), per-band level changes, spectral-shape similarity, issues fixed or introduced, and a null test (aligns and subtracts) that tells whether B is identical, gain-changed, lightly or heavily changed. Use after audio_process to verify an edit did what you intended and nothing else.

Split the audio into distinct sound events (hits, notes, words/syllables, SFX layers) and describe each: start/end, level, attack, tonal vs noisy, brightness, pitch and pitch glide. `as_text` gives one line per event, e.g. "0.52-0.81s | -14dB | tonal | bright | A5 glide -12st | sharp attack". Also reports spacing regularity and the sequence of pitches. The threshold adapts to the background level; set threshold_db (dBFS) to override. split_on_onsets=false keeps continuous sounds (e.g. a sentence) as one event.

Fast metadata without decoding: format, sample rate, channels, duration. Files read through FFmpeg (M4A/AAC, WMA, MP4/MOV/WebM/MKV video) report decoded_via="ffmpeg", the container as format and the codec as subtype; without FFmpeg they fail with install instructions. If `path` is a directory, lists the audio files inside it (recursively).

Transcribe a monophonic line (voice, whistle, solo instrument, synth lead, SFX pitch) into notes with start times, durations, note names and cents deviation, plus a compact text form. Narrow fmin/fmax to the expected range for better accuracy (e.g. bass 40-400, voice 80-1000).

Apply operations in order and write a NEW file (default: <name>_processed.<ext> next to the input; the input is never modified unless output==path and overwrite=true; M4A/video inputs default to WAV output). Each operation is an object, e.g. [{"op":"highpass","hz":80}, {"op":"normalize_lufs","lufs":-16}, {"op":"fade_out","seconds":0.5}]. Ops: gain(db) normalize_peak(db=-1, true_peak=false) normalize_rms(db=-20) normalize_lufs(lufs=-16, ceiling_db=-1, limit=false, measure=integrated|momentary_max) trim_silence(threshold_db=-60, pad_ms=10) denoise(reduction_db=12, sensitivity=1, noise_start, noise_end) [verify speech afterwards with audio_transcribe] limit(ceiling_db=-1, release_ms=60, lookahead_ms=5) compress(threshold_db=-18, ratio=3, attack_ms=10, release_ms=120, knee_db=6, makeup_db=0) fade_in/fade_out(seconds, curve=cosine|linear|equal_power) crop(start, end) pad(start, end) highpass/lowpass(hz, order=4) eq(hz, db, q=1) low_shelf/high_shelf(hz, db) remove_dc to_mono to_stereo resample(rate) reverse polarity(channel=None) loop_crossfade(seconds=0.05) mix(path, at=0, gain_db=0) append(path, gap=0). subtype: e.g. PCM_16, PCM_24, FLOAT, PCM_U8 (8-bit) for WAV. Returns before/after levels and any warnings.

Generate audio from a declarative spec: tones, sweeps, chords, noise bursts, simple melodies and layered sound effects (band-limited oscillators, ADSR envelopes, vibrato, tremolo, filters, panning). Example laser zap: {"layers":[{"wave":"square","freq":1800,"freq_end":200,"duration":0.25, "envelope":{"attack":0.002,"release":0.1},"lowpass":6000}]}. Example melody: {"layers":[{"wave":"triangle","sequence":[["C4",0.25],["E4",0.25],["G4",0.5]]}]}. Full schema: spec = { "sample_rate": 44100, "duration": <seconds, default: end of last layer>, "normalize_peak_db": -1 (null to skip), "layers": [{ "wave": "sine|square|saw|triangle|noise|pink|brown|impulse|dc|silence", # pink/brown are generated circularly, so they loop seamlessly as-is. # impulse: one full-scale sample at the layer start, or a train with "freq" impulses per second. # dc: a constant offset (level set by gain_db; 0 dB = +1.0). silence: exact zeros. "freq": 440 or "A4", "freq_end": <optional sweep target>, "sweep": "exp|linear", "phase": <start phase, degrees>, "start": 0, "duration": <seconds>, "gain_db": 0, "pan": -1..1, "envelope": {"attack": 0.005, "decay": 0, "sustain": 1, "release": 0.01}, "vibrato": {"rate_hz": 5, "depth_cents": 20}, "tremolo": {"rate_hz": 6, "depth": 0.5}, "lowpass": <hz>, "highpass": <hz>, "sequence": [["C4", 0.25], ["rest", 0.25], ["E4", 0.5]] # optional: one note per item, played back to back }] }

How the sound changes over time: per-segment RMS, peak and brightness, plus a loudness sparkline. Streams the file, so it works on recordings of any length; use it to find where to zoom in.

Speech to text (Whisper, runs locally). Returns the text, timed segments, per-word timings with confidence, detected language, speaking rate, long pauses and `uncertain_words` (likely mis-hearings). model: tiny/base/small/medium/large-v3/turbo (bigger = more accurate, slower; first use downloads it). language: ISO code like "en" to skip detection. initial_prompt: names or jargon to help spelling. Use start/end to transcribe part of a long recording.

Render a picture of the sound. kind: 'both' (waveform above spectrogram), 'waveform' or 'spectrogram'. Spectrogram: time left to right, frequency bottom to top, brighter = louder. Horizontal lines are pitched tones/harmonics, vertical stripes are clicks/transients, diffuse haze is noise. Red dashed lines on the waveform mark 0 dBFS (clipping). Zoom with start/end; pitch_overlay draws the detected f0 in cyan. save_to optionally also writes the PNG to disk.