Archive enrichment
Drop 50 hours of unstructured recordings, get back tags, silence and voice/music regions, and a searchable schema. Speaker turns are not part of it: the indexer segments by signal, it does not identify who is speaking.
Audio intelligence Live
SignalLab
A labelled timeline for every recording: speech, music, noise, silence.
Less about creativity, more about structure. SignalLab maps each file into classified regions with confidence, plus hardware-level QA flags (clipping, DC offset, noise floor). The file verdict is a small learned model and a proxy: tested once on 12 October 2026 against 600 real excerpts (MUSAN) plus three environments it had never seen, it found 197 of 200 speech and 190 of 200 music excerpts and 40 of 90 steady-noise ones, and it misses street traffic and the metro. The region timeline comes from a simpler frame classifier.
Live demo · indexer
SignalLab · indexer
SignalLab Indexer
Drop in audio, get back technical tags, semantic hints, risk flags, and a metadata schema your stack can ingest.
Files are decoded in your browser. JSON output is download-only.
How this is measured▸ Drop any audio to fingerprint it: content type, tags and QA flags in one pass.
Use cases
Use cases
QA gates for ingest
Block files at upload time when they fail silence, clipping, or noise-floor thresholds.
Semantic audio search
Find every clip that is speech over a noisy background across a content library without listening to anything twice. The tags describe the signal: voice, music, noise, silence, clipping and the levels behind them.
What makes it different
What makes it different
SignalLab treats audio like data and says how sure it is: every region carries a confidence, the file verdict never claims more than 0.85, and unknown is a real answer. The file verdict is a small learned model and a proxy: tested once on 12 October 2026 against 600 real excerpts (MUSAN) plus three environments it had never seen, it found 197 of 200 speech and 190 of 200 music excerpts and 40 of 90 steady-noise ones, it misses street traffic and the metro, and its confidence drops where it is wrong but is not a probability. The region timeline comes from a simpler frame classifier. Schema-first, API-first, and built so the people downstream (editors, ML teams, speech pipelines) can actually use the output.
Also in SignalLab
- Silence/clipping detector
Catch the bad parts before your editor does.
- Speaker/event marker (concept)
Auto-mark turns, silences and scene changes. Silence regions exist today; speaker turns do not.
- Sound recognition (concept)
Say what is in a recording, from broad to fine: animal, bird, vehicle, a keyboard. Not built yet. SignalLab labels voice, music, noise, silence and clipping today.
Roadmap
Roadmap
Now 01
- Speech-calibrated region classifier (voice/music/noise/silence + confidence)
- Clipping, DC-offset and noise-floor QA flags
- Timeline UI with click-to-play regions
Next 02
- Batch indexing for archives
- Long-file chunking
- Music-under-speech detection (the TV in the background)
- REST API + webhooks
Later 03
- ASR hallucination guard (text where no speech was)
- Fleet health reporting per device
- On-prem option
- Custom calibration per recogniser
Read the system, not the marketing.
Free in the browser · API + MCP at scale
Run SignalLab from your own code or any AI agent.
The demo above runs free, on your device.The lab above runs free, on your device. The same analysis is a hosted REST API and an MCP server, one key. Loudness is verified against EBU R128 / BS.1770-4. SignalLab exposes 1 endpoint / 1 MCP tool.
REST · POST
-
/v1/signallabcontent tagging + metadata
MCP tools · https://audiolab.tools/mcp
- index_signal
Quick start
curl -X POST https://audiolab.tools/v1/signallab \
-H "Authorization: Bearer $AUDIOLAB_KEY" \
-H "content-type: application/json" \
-d '{"url":"https://your-audio.wav"}'Where this leads.
Continue the workflow
MixLab
Master a region
Send the full file to MixLab for BS.1770-4 LUFS analysis, true-peak, and tonal balance on the program feed.
VoiceLab
QA the voice regions
When SignalLab finds voice content, push the same audio to VoiceLab for sibilance and noise-floor scoring.
HearLab
Recognise the scene live
Where SignalLab labels a recording, HearLab names the scene live from your microphone: quiet, speech, babble, traffic or wind.
CueLab
Build a cue sheet from regions
Indexed regions translate naturally into broadcast cue triggers. Plan the sequence in CueLab.
SkillLab
Train the classifier in your head
SkillLab includes drills on voice-vs-music-vs-noise classification, the same task SignalLab automates.
MixLab, VoiceLab, and SignalLab hand audio off to one another without re-uploading. When you finish an analysis in one of those, the → Lab buttons in the demo chrome carry your decoded audio directly into the next lab’s analysis flow. The other labs open fresh.