Local-first AI audio editor · v0.2.0

Describe it.Get pro-grade audio edits.

Twenty-two minutes of raw tape. Two speakers, dead air, music that fights the voice. Watch it become an episode.

You say

Cut the dead air, duck the music under my voice, and render it.

No menu hunting. No preset chain. A sentence.

It works

loadtranscribetruncate_silenceset_clip_enveloperender_final

Three stretches of silence gone. The tail closes up behind them — one undoable step, not a destructive render.

And the music gets out of the way

Keyed on the transcript, not on level — so a breath does not trigger it and a quiet line does not escape it. The result is an ordinary automation curve you can drag.

Rendered.

Every step is a node you can undo, branch, or export as a recipe and run over the next twelve episodes.

Download for Mac

Download for Windows

Unsigned dev builds · Mac (universal) · Windows 10/11 · Linux

0: audio uploaded to any server
6: Anthropic · OpenAI · Gemini · Groq · OpenRouter · Ollama
93: the agent can call, from fade to beat-warp
100%: DSP runs locally, always

The gap

Pro DAWs are powerful but take 100+ hours to learn. AI tools are easy but shallow.

Nobody offers conversational, multi-track production at professional DSP quality. edytlab is the agent layer that plans, executes, and iterates over a real audio engine — stem separation, transcription, time and pitch, a full effect chain — in a session you can actually trust and steer.

How it stacks up

Pro quality. Zero friction.

The column that matters is the one with every row filled.

Feature	Traditional DAW	AI Audio Tools	edytlab
Learning curve	100+ hours	Minutes	Minutes
Audio quality
Privacy — local audio
Natural language input
Session branching
BYO LLM key
Stem separation
MCP extensibility

See it in action

A walkthrough of the conversational editing workflow.

edytlab

a3f1

0:00─2:34

vocals.mp3

drums.mp3

bass.mp3

Agent

Built for producers who want help — not handcuffs.

What makes edytlab different from cleanup tools, preset chains, and shallow AI wrappers.

Edit the words, not the waveform

Transcribe a take and the transcript becomes the editor. Select a sentence to select the audio; delete it and the audio is cut and the gap closed. The remaining word timings shift to match, so the transcript keeps lining up with what you can hear. It is an ordinary edit underneath — undoable, branchable, and recorded like any other.

Conversational multi-track

Mash A's vocals over B's drums, key-match, give me three takes. The agent plans, executes multi-track mixing, and renders branches — all from a single prompt.

Pro-grade DSP

Pure Rust audio graph (cpal · symphonia · rubato · realfft) with Demucs stem separation and Whisper transcription. Time-stretch, pitch-shift and formant preservation run on a phase vocoder written for this project — no C dependency in the audio path.

Local-first

Your audio never leaves your machine. The DSP engine runs on-device; only chat tokens hit your chosen LLM provider.

Undo, branch & compare

Every state is a DAG node. Ctrl+Z/Y traverse the branch history. Fork, A/B compare, and revert are first-class — not hidden behind a linear undo stack.

Bring your own LLM

Anthropic, OpenAI, Gemini, Groq, OpenRouter or a local Ollama daemon. Keys live in your OS keychain. Per-model agent profiles tune tools and behaviour. Swap providers without reinstalling.

Time, pitch and timing

Stretch without moving the pitch, shift pitch without moving the clock, and preserve formants so a shifted voice still sounds like the same person. Warp a performance onto a beat grid in a single pass — no seam at the beats.

Export that plays anywhere

WAV when you want the samples, FLAC for lossless at about half the size, MP3 when it has to open on anything. Loudness-normalise to a LUFS target — the number streaming platforms actually use — with a true-peak ceiling so it never clips getting there.

MCP extensibility

Register Model Context Protocol servers from Settings to give the agent new tools. Wire in stdio JSON-RPC servers and extend what edytlab can do without touching core code.

The interface

Talk to it — or reach in and move things yourself.

The agent is the fast path, not the only one. Faders, automation curves and clips are all directly editable, and every change lands in the same undoable session graph the agent writes to.

mixer

drums

-4.5 dBC

bass

-2.0 dBL35

vocal

+1.5 dBR20

mutesolo

Gain, pan, mute and solo per track — by hand, without asking the agent. Every move is one undoable step in the session graph.

automation

Draw a volume curve on the clip. Click to add a point, drag to move it, arrows to nudge. The render interpolates between them per frame.

clips

take.wav

Cut a track and the seam is visible. Select a clip, drag it later, delete it — the waveform and the arrangement stay in step.

The toolbox

93 tools the agent can reach for.

Named, not counted. Ask in plain language and the agent picks — but if something is not on this list, it cannot do it, and it will tell you so rather than pretend.

gainnormalizenormalize_loudnesslevelercompressorlimiternoise_gateeqlow_pass_filterhigh_pass_filternotch_filternoise_reductionclick_removalde_esservocal_reductiontruncate_silencereverbechodistortionphasertremolostereo_wideneradd_effectset_effect_paramsset_effect_bypassedreorder_effectsremove_effecttime_stretchpitch_shiftchange_speedalign_to_beatcut_rangetrimsplit_clipmove_clipremove_clipcopy_regionpaste_regioninsert_silencerepeat_selectiontime_shiftset_clip_envelopeset_track_gainset_panmute_tracksolo_trackcreate_busset_sendmix_to_new_trackanalyze_trackstorage_reportplot_spectrumsilence_findertranscribeseparate_stemsfork_nodecompare_nodesrevert_toapply_diffname_noderender_previewrender_finalexport_multipleexport_labels

Level

gain
normalize
normalize_loudness
leveler
compressor
limiter
noise_gate

Tone

eq
low_pass_filter
high_pass_filter
notch_filter

Repair

noise_reduction
click_removal
de_esser
vocal_reduction
truncate_silence

Effects

reverb
echo
distortion
phaser
tremolo
stereo_widener

Effect chains

add_effect
set_effect_params
set_effect_bypassed
reorder_effects
remove_effect

Time & pitch

time_stretch
pitch_shift
change_speed
align_to_beat

Arrangement

cut_range
trim
split_clip
move_clip
remove_clip
copy_region
paste_region
insert_silence
repeat_selection
time_shift
set_clip_envelope

Mixing

set_track_gain
set_pan
mute_track
solo_track
create_bus
set_send
mix_to_new_track

Analysis

analyze_track
storage_report
plot_spectrum
silence_finder
transcribe
separate_stems

History

fork_node
compare_nodes
revert_to
apply_diff
name_node

Export

render_preview
render_final
export_multiple
export_labels

Plus track and format management — add, remove, rename and duplicate tracks, resample, mono/stereo conversion, tone and noise generators, markers and label import/export.

How it works

Three steps. No DAW manual required.

1. Drop audio

Drag in songs, stems, or a folder of takes. WAV, MP3, FLAC, OGG — decoded with symphonia, no upload, no waiting.

2. Talk to the agent

“Mashup A over B, key-match, give me 3 takes on the drop.” The agent plans, shows the plan, and renders branches you can A/B.

3. Export

Pick the branch you like. Export to WAV, FLAC for a lossless file about half the size, or MP3 when it has to play anywhere. The session graph keeps every alternative.