Overcast Wiretap

overcast-wiretap

Use this skill to analyze audio recordings you already hold (a call, a voicemail, a field recording): how many people speak, what the background reveals about where it was recorded, and whether two clips share a voice or a phrase. Use the broad overcast skill and overcast/reference/verbs.md for exact flags.

Workflow

1. Transcribe the recording and read its background scene on the **default

backend**. --describe surfaces the whole audio scene (traffic, trains, a PA announcement, church bells) — the "enhance the background noise" move that places a recording — and is a tinycloud/Cloudglue multimodal feature, so do all the transcript/describe work FIRST, before binding any speech-only provider:

overcast doctor --json
overcast case init --json
overcast listen ./call.wav --json                 # speech transcript + time-anchored segments
overcast listen ./call.wav --describe --json       # background audio scene (tinycloud describe)
overcast view ./call.wav --spectrogram --json      # visual inspection artifact (tones, hums, edits)

1. Isolate voices from noise and re-transcribe the cleaned track — a second pass

often recovers words the first missed:

overcast enhance ./call.wav --ops voice-isolate,denoise --json   # bundled-ffmpeg filter (fast, lossy)
overcast enhance ./call.wav --ops separate --json                # true source separation → a track per speaker
                                                                 # (local pyannote: scripts/visual-db-uv.sh --voice + HF_TOKEN, or fal)
overcast listen  --json

1. Separate the speakers. --diarize needs a diarize-capable listen provider

(the default tinycloud/Cloudglue listen is speech-transcript only and rejects --diarize). Bind ElevenLabs for this — but do it LAST, after the describe / multimodal steps above, because ElevenLabs Scribe is speech-only and drops the audio-scene describe:

overcast provider setup apply --verb listen --choice elevenlabs --yes --json
overcast listen ./call.wav --diarize --json        # ->  (speaker-labeled)

1. Verify a speaker's identity across recordings with the local voice-print DB

(speaker embeddings, not phrase matching — scripts/visual-db-uv.sh --voice, no token needed). Scores are rank scores, not liveness: a cloned/synthetic voice can score high, so corroborate before naming anyone:

overcast index create voices --type voice-print --local --json
overcast voice add ./call.wav --index voices --json               # enroll each recording
overcast voice match ./voicemail.m4a --index voices --json        # which recordings share this voice?
overcast voice match ./call.wav ./known-sample.wav --diarize --json  # WHICH diarized speaker matches (HF_TOKEN)

voice answers WHO is speaking (enroll a reference, then rank where/which). To ask whether two clips are the SAME RECORDING (a re-upload/leak of an identical file — a different question), fingerprint them with the local audio-fp DB, which matches exact audio through transcode/noise but NOT pitch/speed:

overcast index create clips --type audio-fp --local --json
overcast audio add ./call.wav --index clips --json
overcast audio match ./leaked.mp3 --index clips --min-margin 2 --json   # same recording? time-offset aligned

Full drills: overcast-voiceprint (WHO is speaking) and overcast-audio-match (same recording surfaced again).

1. Record per-speaker and per-clue observations, then correlate across recordings.

Cite the speaker-labeled `` for who-said-what claims (not the step-1 transcript record):

overcast note "Speaker 2: PA announces 'platform 4' at 00:38 → rail station" --ref  --at 38 --confidence medium --json
overcast ask "which recordings share a speaker, phrase, or background cue? cite record.id + time" --verb listen --json

1. Turn confirmed clues into findings and export; always leave a tldr note:

overcast finding create "call.wav and voicemail.m4a share Speaker 2's phrasing + station PA — likely same caller/location" --ref  --confidence medium --json
overcast note "3 recordings; 2 speakers on call.wav; background = rail station; cross-clip voice overlap on 2 of 3" --tag tldr --json
# Wait for the note result before exporting, so the TL;DR is included.
overcast brief --export ./wiretap.html --json

Output

A per-recording speaker breakdown with timestamps, the background-scene clues that locate or date it, any cross-clip voice/phrase overlaps, and the spectrogram artifacts — each cited by record.id + media.at.

Caveats

--diarize needs a diarize-capable listen provider: the default tinycloud/Cloudglue listen transcribes speech only and errors on --diarize — bind ElevenLabs (provider setup apply --verb listen --choice elevenlabs --yes) for speaker separation. Bind it LAST: ElevenLabs Scribe is speech-only and drops the audio-scene --describe, so run all transcript/describe/multimodal work on the default backend first, then rebind for the diarize pass. Diarization LABELS speakers ("Speaker 1/2"), it does not IDENTIFY them — a name is a corroborated inference, never a diarizer output. voice match scores speaker similarity, not liveness — a cloned/synthetic voice can score high and the same speaker scores lower across languages or heavy compression, so treat a voice match as a lead to corroborate, never an identification. voice-isolate on the bundled ffmpeg is a filter, not source separation — for a real per-speaker split use enhance --ops separate (local pyannote via scripts/visual-db-uv.sh --voice + HF_TOKEN and the accepted pyannote license, or fal), or bind ElevenLabs (--verb enhance --choice elevenlabs) for stronger isolation; either way re-listen to confirm the cleaned transcript rather than trusting it blind. Background-cue geolocation is suggestive, not definitive.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: kdr
• Source: kdr/overcast
• License: Apache-2.0
• Homepage: https://overcast.video

Install and usage instructions live in the source repository linked above.