Audio enhancement

Rakkr can produce a noise-suppressed, level-normalized enhanced rendition of a recording alongside the untouched raw master. Enhancement runs entirely in-process on the recorder agent — no extra packages are deployed to nodes — using two pure-Rust engines with embedded models:

DeepFilterNet3 (default) — a deep-learning denoiser, strongest on non-stationary noise (HVAC, crowd, handling).
RNNoise — a lighter recurrent-network denoiser.

Both operate at 48 kHz. On a typical recorder node DeepFilterNet3 runs at roughly 12× real time and RNNoise far faster, so enhancement is comfortably cheaper than real time.

Raw is always preserved

Enhancement is an additional, switchable layer — never a replacement. The raw capture is kept as the master so you can always reproduce exactly what the microphones heard. DNN denoisers can occasionally alter content (suppress quiet speech, smear overlapping talkers), so for a system of record the raw rendition is the source of truth and the enhanced rendition is a convenience for intelligibility.

This is a data-integrity choice, not a legal one: applying a filter does not make a recording inadmissible.

The chain (configured per recording profile)

The enhancement chain lives on the recording profile (the preset/template), so it is RBAC-gated and audited like any other settings change. Every stage is independently toggleable with configurable parameters, applied in this order:

Denoise — DeepFilterNet3 or RNNoise, applied in-process before any ffmpeg filter (default on, DeepFilterNet3).
High-pass — remove rumble / HVAC / handling (default on, 80 Hz).
Low-pass — optional high-frequency roll-off (default off).
De-esser — tame sibilance (default off).
Compressor — even out speakers at different mic distances (default off).
Loudness normalization — EBU R128, so every recording sits at a consistent level (default on, −16 LUFS / −1.5 dBTP / 11 LRA).
Noise gate — optional, threshold in dB (default off).

keepRaw (default on) controls whether the raw master is uploaded alongside the enhanced rendition. Edit all of this under Settings → recording profile.

v1 targets voice profiles (mono / stereo / mono-to-stereo-mix); the denoiser runs on the channel-mapped mono signal.

Playback: raw vs enhanced

When a recording has both renditions, the playback dock shows an Enhanced / Raw toggle. Default playback is the enhanced rendition. The renditions are served by:

GET /api/v1/recordings/:id/stream                  # default (enhanced if present)
GET /api/v1/recordings/:id/stream?rendition=raw    # raw master
GET /api/v1/recordings/:id/stream?rendition=enhanced

Live listen: raw vs enhanced

The live listen monitor has the same Raw / Enhanced toggle. Switching it restarts the monitor session with the chosen rendition; the controller serves the matching chunk and the recorder agent only spends CPU denoising the live stream while a listener is actually requesting enhanced audio (on-demand). When no one is listening enhanced, the agent produces only the raw monitor chunk.

Because the agent learns of the request on its next controller config poll, enhanced live audio appears a few seconds after you flip the toggle; until the first enhanced chunk arrives the stream falls back to raw. Live enhancement always uses DeepFilterNet3 and needs no per-node configuration — it is driven entirely by the listener toggle.

How it is produced

For each recording the agent renders the raw rendition as before, then — when the profile enables enhancement — produces the enhanced rendition: ffmpeg channel-maps and downmixes the raw capture to 48 kHz mono, the in-process denoiser cleans it, and ffmpeg applies the enabled voice-chain filters and encodes to the profile codec. The enhanced rendition is uploaded as the primary (?rendition=enhanced) and the raw render as a supplementary master (?rendition=raw). Enhancement is best-effort: if it fails, the recording still completes with the raw rendition.