Audiolla

Remove broadband background noise (ML path via UVR MelBand Roformer). Thin shim — prefer ``restore`` with engine=uvr-denoise, or ``noise_reduce`` with engine=uvr-denoise for the same result. engine default: uvr-denoise.

Get the full step list of a preset before running it.

Detect digital clipping in an audio file. Returns: clipped, clip_count, clip_ratio, peak_db, duration_sec, sample_rate, channels.

Speaker diarization — who spoke when. Returns time-stamped speaker segments. engine options: pyannote (default, requires HUGGINGFACE_TOKEN).

Analyse a track for DJ use: BPM, musical key, Camelot wheel position, and integrated LUFS loudness. Requires librosa-analyze + chord-detect.

Generate a MIDI drum pattern from a step-sequencer spec. pattern: dict of instrument → step array (1=hit, 0=rest; or 1-127 for velocity). Instruments: kick, snare, hihat, hihat_open, ride, crash, clap, rim, cowbell, etc. Returns MIDI base64 or uploads to output_url.

512-dim L2-normalised audio embedding via LAION CLAP. With query_text, also returns cosine similarity to the text description. Requires HF model cache.

Neural speech and vocal enhancement via DeepFilterNet DF3. Provide exactly one of `file_path` or `file_url`. Writes audio to `output_path` xor `output_url`; response is `{path,size,engine,output_format}` or `{url,size,engine,output_format}`.

Parametric EQ. bands: [{freq, gain_db, width_hz (opt)}]. Writes audio to `output_path` xor `output_url`; response `{path,size,...}` or `{url,size,...}`.

Apply fade-in/fade-out. curve: tri/qsin/esin/hsin/log/exp/lin/etc. At least one of fade_in/fade_out must be > 0. Writes audio to `output_path` xor `output_url`; response `{path,size,...}` or `{url,size,...}`.

Find the best seamless loop point in audio using beat-grid MFCC similarity. Returns loop_start_sec, loop_end_sec, bars, score, tempo_bpm, and top candidates.

Chromaprint audio fingerprint via fpcalc. Returns ``{duration, fingerprint}`` (+ ``fingerprint_raw`` int array if ``return_raw=True``).

Generic effects chain. Provide exactly one of `file_path` or `file_url`. `effects` is an ordered list of {type, params} — type names match pedalboard classes (Compressor, Reverb, Chorus, Delay, PitchShift, HighShelfFilter, ...). Writes audio to `output_path` xor `output_url`.

Text-to-music generation. ``engine`` selects the generator: - ``ace-step`` — Apache 2.0, ~4 min vocal+instrumental songs. - ``diffrhythm`` — Apache 2.0, ~4m45s songs, lowest VRAM. - ``stable-audio-open`` — Stability Community License, 47s cap, loops / SFX / textures (no vocals). ``lyrics`` applies to ace-step and diffrhythm only; ignored by stable-audio-open. Writes audio to `output_path` xor `output_url`; ``output_path`` to stage in FILES_DIR or ``output_url`` for a presigned PUT. Long generations may take multiple minutes — use the REST endpoint with ``async_job=true`` for fire-and-forge…

Read a staged file (base64-encoded back).

Get the status and result of an async job by ID.

Harmonic/percussive source separation via librosa HPSS median filter. Provide exactly one of `file_path` or `file_url`. Returns ``{stems: {harmonic: <path|url>, percussive: <path|url>}, output_format}`` — pass ``output_paths={'harmonic':..., 'percussive':...}`` xor ``output_urls={...}``. ``margin`` ≥1.0 controls separation aggressiveness; ``kernel_size`` sets the median filter width (odd int, default 31).

Probe audio file metadata — duration, sample_rate, channels, codec, bit_depth, format, size_bytes. No engine required; works on any format.

Detect source key then pitch-shift to target_key (e.g. C, F#, Bb). Writes audio to `output_path` xor `output_url`; response `{path,size,...}` or `{url,size,...}` plus {source_key, target_key, semitones}. Requires chord-detect + stretch engines.

List configured engines + their capabilities.

List files in the staging area.

List async jobs submitted to the job queue. status: filter by 'pending', 'running', 'completed', 'failed', 'cancelled'.

List pipeline op slugs available in ``/v1/pipeline`` and presets.

List server-side curated workflows (name + description). Use ``run_preset`` to execute one against a file.

Repeat audio count times (minimum 2). Writes audio to `output_path` xor `output_url`; response `{path,size,...}` or `{url,size,...}`.

Measure integrated loudness (LUFS) via pyloudnorm. Returns JSON only. Use ``normalize`` to produce loudness-normalized audio.

Compute the RMS loudness envelope of an audio file as a time series. Returns curve (list of {time_sec, rms_db}), duration, sample_rate, points.

Master audio. Provide exactly one of `file_path` or `file_url`. For `mode=reference`, also provide one of `reference_path` or `reference_url`. For `mode=chain`, set `preset` (transparent or loud). Writes audio to `output_path` xor `output_url`; response `{path,size,...}` or `{url,size,...}`. target_lufs is optional in both modes.

Monophonic pitch tracking via pyin. Returns the pitch contour and (with ``as_midi=True``) a MIDI file of quantised (writes to ``output_path`` xor ``output_url``) notes.

Encode stereo audio to Mid/Side or decode M/S back to L/R. mode: 'encode' (L+R→Mid, L-R→Side) or 'decode' (Mid+Side→L/R). Returns audio base64 or uploads to output_url.

Compose MIDI bytes from a song spec. `spec` is the same JSON shape accepted by /v1/midi/compose: ``{tempo_bpm, time_signature, tracks: [{program, channel, notes: [{pitch, start_beats, duration_beats, velocity}]}]}``. Writes MIDI to `output_path` xor `output_url`; response carries `size` and either `path` or `url`.

Compose AND render a song spec in one call — JSON in, audio out. Convenience wrapper around midi_compose + midi_render.

Add random timing jitter and velocity variation to MIDI notes. timing_ms: max ±offset per note in ms. velocity_pct: max ±velocity change as % of 127. seed: optional RNG seed for reproducibility. Returns MIDI base64.

Parse a Standard MIDI File and return JSON describing tempo, time signature, key signature, and per-track stats. Counterpart to midi_compose.

Snap MIDI note timings to the nearest rhythmic grid. grid_beats: 0.25=16th note, 0.5=8th, 1.0=quarter. Writes MIDI to `output_path` xor `output_url`; response `{path,size,...}` or `{url,size,...}`.

Render a MIDI file to audio via fluidsynth + SoundFont. Provide exactly one of `file_path` (staged MIDI) or `file_url` (remote MIDI — subject to fetch policy). `soundfont_path` optionally overrides the server's default SoundFont with a staged ``.sf2``. Writes audio to `output_path` xor `output_url`; response `{path,size,...}` or `{url,size,...}`.

Transform a MIDI file: transpose / quantize / change tempo / filter channels. Writes MIDI to `output_path` xor `output_url`; response `{path,size,...}` or `{url,size,...}`.

Mix multiple audio tracks with per-track gain. tracks: list of {file_path or file_url, gain_db (optional, default 0.0)}. Requires at least 2 tracks. Writes audio to `output_path` xor `output_url`; response `{path,size,...}` or `{url,size,...}`.

Multiband compression. ``crossovers_hz`` is an ascending list of crossover frequencies in Hz (length N). ``bands`` is a list of N+1 compressor specs, each with ``threshold_db``, ``ratio``, and optional ``attack_ms``, ``release_ms``, ``makeup_db``. Splits signal into N+1 bands with zero-phase LR4-equivalent filters, compresses each independently, sums the result.

Noise reduction — DSP or ML depending on engine. engine=noise-reduce (default): DSP via noisereduce, no GPU. stationary=True targets constant noise (hum/hiss); False = adaptive. prop_decrease [0,1] scales reduction strength. engine=uvr-denoise: ML via UVR MelBand Roformer (SDR 28), higher quality. stationary and prop_decrease are ignored for ML engine.

Normalize audio to a target LUFS level via pyloudnorm. Provide exactly one of `file_path` or `file_url`. Writes audio to `output_path` xor `output_url`; response carries ``measured_lufs`` and ``target_lufs`` plus `{path,size,...}` or `{url,size,...}`.

Onset (transient) detection. Returns time + relative strength for each detected attack.

Pan audio in the stereo field. position: -1.0=hard left, 0.0=center, 1.0=hard right. Writes audio to `output_path` xor `output_url`; response `{path,size,...}` or `{url,size,...}`.

Pitch-correct audio toward the nearest chromatic semitone. Uses pyin F0 detection. strength=1.0 is full correction, 0.0 is bypass. Returns corrected audio as base64 or uploads to output_url.

Upload a file (base64-encoded) to the staging area.

Repair audio artifacts: interpolate clipped samples and/or remove mains hum. declip: fix digital clipping (interpolates ±0.999 saturation). dehum: notch-filter 50 or 60 Hz hum + harmonics. hum_freq: 50 (EU) or 60 (US) Hz.

Remove reverb, echo, or noise from audio using UVR models. engine=uvr-dereverb (default): BS-Roformer reverb removal. engine=uvr-deecho: VR Architecture echo removal. Set aggressive=True for harder echo removal. engine=uvr-denoise: MelBand Roformer noise removal (SDR 28). For DSP-based noise reduction use noise_reduce with engine=noise-reduce. Writes audio to `output_path` xor `output_url`; response is `{path,size,engine,output_format}` or `{url,size,engine,output_format}`.

Reverse audio playback direction. Writes audio to `output_path` xor `output_url`; response `{path,size,...}` or `{url,size,...}`.

Run an ad-hoc op pipeline: ``steps=[{op, params}, ...]``. Each step's output feeds the next. See ``list_ops`` for available op slugs.

Run a curated preset pipeline against an input file. Writes audio to `output_path` xor `output_url`; response `{path,size,steps}` or `{url,size,steps}`.

Music structure segmentation via recurrence-matrix clustering. Returns N labelled ranges; structurally similar regions share a label so verse/chorus repetition is easy to spot.

Demucs stem separation. Provide exactly one of `file_path` or `file_url`. Three output modes (mutually exclusive — pass at most one): - default: per-stem audio paths under `stems` (one path per stem). - `output_paths={stem_name: path}`: stage each stem in FILES_DIR; response has `staged_stems` mapping stem -> {path, size}. - `output_urls={stem_name: presigned_put_url}`: PUT each stem to its URL; response has `uploaded_stems` mapping stem -> {url, size}. Both per-stem maps must cover every returned stem (no partials).

Duck primary audio when trigger audio is loud. Provide primary via file_path/file_url, trigger via trigger_file_path/trigger_file_url. Writes audio to `output_path` xor `output_url`; response `{path,size,...}` or `{url,size,...}`.

Find silent ranges + optionally auto-trim. ``trim_mode='edges'`` removes leading/trailing silence; ``'all'`` removes every detected silence. When `trim_mode` is set, writes trimmed audio to `output_path` xor `output_url`.

Cosine similarity between two audio files via CLAP embeddings. Returns {similarity: float [-1,1], dim: 512}. Requires clap-embed model cache.

Slice audio at beat positions detected by librosa-analyze. Returns a base64-encoded ZIP of numbered beat slice WAV/MP3 files.

Change playback speed without pitch shift via ffmpeg atempo. speed=0.5 halves speed; speed=2.0 doubles. Range: 0.1–10.0. Writes audio to `output_path` xor `output_url`; response `{path,size,...}` or `{url,size,...}`.

Split audio into segments. mode=equal: requires count>=2, splits into equal time parts. mode=silence: splits on quiet gaps (uses threshold_db/min_duration_sec). DEPRECATED in v1.0.0 — raises ValueError; use REST `POST /v1/audio/split` with `output_path`.

Analyse the stereo field: L/R correlation, width, balance, mono compatibility. Returns correlation (-1..1), width (0=mono, 1=normal), balance_db, phase_issues flag.

Adjust stereo image width via M/S processing. width=0.0 → mono, 1.0 → original, >1.0 → wider. Range: [0.0, 3.0]. Writes audio to `output_path` xor `output_url`; response `{path,size,...}` or `{url,size,...}`.

Time-stretch and/or pitch-shift audio. tempo_factor=0.5 halves speed; pitch_semitones=12 shifts one octave up. Writes audio to `output_path` xor `output_url`.

Audio tagging via Audio Spectrogram Transformer — returns top-K AudioSet class labels with confidence scores. Requires HF model cache.

pysox DSP transform chain. Provide exactly one of `file_path` or `file_url`. `operations` is a list of {op, params} — valid ops: gain, equalizer, compand, reverb, pitch, tempo, rate, channels, trim, pad. Writes audio to `output_path` xor `output_url`.

Shape transients via dual-compressor attack/sustain blending. attack_gain_db > 0 makes drums punchier; sustain_gain_db < 0 cuts room tail. Returns processed audio base64 or uploads to output_url.

Cut audio to [start_sec, end_sec). end_sec required and must be > start_sec. Writes audio to `output_path` xor `output_url`; response `{path,size,...}` or `{url,size,...}`.

Voice activity detection via silero-vad. Returns speech/non-speech segments with timestamps and overall speech ratio.

Render audio visualization. ``engine`` selects the visualization engine (default: ``ffmpeg-render``). ``mode`` controls the output type: - ``spectrogram`` — static PNG via ffmpeg showspectrumpic (color + scale apply) - ``waveform`` — static PNG via ffmpeg showwavespic (color applies, no scale/fps/container) - ``spectrum`` / ``waves`` / ``cqt`` / ``freqs`` / ``volume`` / ``vectorscope`` / ``phasemeter`` / ``histogram`` — animated MP4/WebM video Output: PNG/MP4/WebM bytes are written to ``output_path`` xor ``output_url``; response is ``{path, size, ...}`` or ``{url, size, ...}``. Pass ``out…