DSP Pipeline & Effects Reference

Echidna's audio processing lives in libech_dsp.so (the DSP engine) and is invoked from libechidna.so inside each hooked audio process (on a device where the hook is active — device-gated; see Verification). Captured audio blocks are processed synchronously inside the hooked read callback to minimise copies and latency; a hybrid mode can offload heavier work to a worker thread when a preset asks for quality over latency.

This page documents the processing chain, every effect stage and its parameters (as declared in native/dsp/), the latency/quality modes, the built-in preset catalog, and the signed plugin system.

---

The processing chain

Effects run in a fixed order. The dry input is preserved and blended back at the end by the mix bus, so Dry/Wet governs how much processed signal you actually hear.

input ──► Gate ──► EQ ──► Compressor/AGC ──► Pitch ──► Formant ──► Auto-Tune ──► Reverb ──► Plugins ──► Mix (dry/wet + output gain) ──► output │ ▲ └────────────────────────────────── dry signal ─────────────────────────────────────────────────────────┘

This is the real order in the engine: signed plugin effects are inserted immediately before the mix bus, so they receive the fully conditioned wet signal. Each stage has an independent on/off toggle; a disabled stage passes audio through untouched.

The C entry points (native/dsp/include/echidna/dsp/api.h) are:

• ech_dsp_initialize(sample_rate, channels, quality_mode)
• ech_dsp_update_config(json_config, json_length) — apply a preset (JSON validated against safe ranges before it takes effect)
• ech_dsp_process_block(input, output, frames) — process one interleaved float block
• ech_dsp_shutdown()

---

Latency & quality modes

Two orthogonal knobs control the latency/quality trade-off.

Quality mode (ech_dsp_quality_mode_t, chosen at init and per preset):

Mode	Target end-to-end latency	Notes
Low-Latency	10–15 ms	Cheap transforms only; forces the pitch shifter to its granular backend.
Balanced	~20 ms
High-Quality	30 ms+	Enables the phase-vocoder pitch backend and heavier stages.

Processing mode (per preset):

• Synchronous — the block is processed in-callback. Lowest latency; the default (block_ms defaults to 15).
• Hybrid — the callback copies to a lock-free ring buffer and a worker thread applies heavy transforms, returning the latest available output. Higher latency, higher quality for non-real-time uses.

Latency-mode safety

Under Low-Latency, the engine overrides a preset's request for the high-quality pitch backend and uses the granular one instead, keeping the in-callback cost bounded. Presets tagged LL are tuned to stay inside the low-latency budget.

---

Effect stages & parameters

Ranges and defaults below are taken from the effect headers in native/dsp/src/effects/ and the parameter reference in spec.md. "Default" is the value the parameter struct ships with.

The Effects tab exposes each stage with an independent on/off toggle. A disabled stage passes audio through untouched. You can preview any preset's chain locally in the Lab — no root required.

1. Noise Gate

Attenuates signal below a threshold with soft attack/release and hysteresis to avoid chatter. The gate opens only after the smoothed envelope reaches threshold + hysteresis and closes after it falls to threshold - hysteresis. Envelope, open/closed state, and gain persist across callbacks, so release continues to silence instead of being reset at each block boundary.

Parameter	Range	Default	Unit
Threshold	−80 … −20	−45	dBFS
Attack	1 … 50	5	ms
Release	20 … 500	80	ms
Hysteresis	0 … 12	3	dB

2. Parametric EQ

A stack of biquad bands; select 3 / 5 / 8 bands. Each band is an independent peaking filter.

Per-band parameter	Range	Default	Unit
Frequency	20 … 12 000	1000	Hz
Gain	−12 … +12	0	dB
Q	0.3 … 10	1.0	—

Named starting points: Phone, AM Radio, Warmth, Bright, De-mud, De-harsh, V-curve. Keep |gain| ≤ 6 dB per band for natural-sounding presets.

3. Compressor / AGC

Dynamic range control with a manual mode and an automatic (AGC) mode.

Parameter	Range / values	Default	Unit
Mode	Manual / Auto (AGC)	Manual	—
Threshold	−60 … −5	−24	dBFS
Ratio	1.2:1 … 6:1	3:1	—
Knee	Hard / Soft (soft amount 0 … 12)	Hard	dB
Attack	1 … 50	5	ms
Release	20 … 500	120	ms
Makeup gain	0 … +12 (auto in AGC)	0	dB

Ratios above 4:1 with a fast release can pump — the UI hints at this.

4. Pitch Shift

Shifts pitch with a granular (low-latency) or phase-vocoder (high-quality) backend.

Parameter	Range / values	Default	Unit
Semitones	−12 … +12	0	st
Fine	−100 … +100	0	cents
Quality	LL (granular) / HQ (phase vocoder)	LL	—
Preserve formants	On / Off	Off	—

Safe range is −6 … +6 st (a warning is shown beyond ±8); Low-Latency prefers ≤ ±4. When "preserve formants" is on, formant correction is delegated to the formant stage.

The Pitch stage. The UI warns beyond the safe range (±8 st) and, under Low-Latency, forces the granular backend to keep in-callback cost bounded.

5. Formant Shift

Alters perceived vocal-tract size independently of pitch. The Pitch Shift screenshot above also shows the formant-preserve control that ties these two stages together.

Parameter	Range	Default	Unit
Cents	−600 … +600	0	cents
Intelligibility assist	On / Off	Off	—

Safe range is −300 … +300 cents (warn beyond ±450). Intelligibility assist tilts the EQ slightly to preserve clarity at large shifts.

6. Auto-Tune (pitch correction)

A monophonic pitch detector quantises to the nearest note of the selected key/scale, then pitch-shifts to the target and blends by snap strength.

Parameter	Range / values	Default	Unit
Key	C, C♯/D♭, D … B	C	—
Scale	Major, Minor, Chromatic, Dorian, Phrygian, Lydian, Mixolydian, Aeolian, Locrian	Chromatic	—
Retune speed	1 … 200	50	ms
Humanize	0 … 100	50	%
Flex-Tune	0 … 100	0	%
Formant preserve	On / Off	Off	—
Snap strength	0 … 100	100	%

Retune under ~10 ms sounds robotic (the "FX" region); slow retune is more natural. The Diagnostics Tuner View shows detected vs. target note in real time.

Detection uses a rolling, per-channel analysis history and updates on a sample-rate-derived hop, so 64–2048-frame callbacks do not need to contain a complete pitch period. Correction is applied by a stateful realtime pitch shifter whose delay/phase history survives block boundaries. Retune smoothing uses the callback's frame duration, keeping the control response consistent at 44.1, 48, and 96 kHz. Silence, low-correlation input, invalid buffers, and callbacks larger than prepared realtime capacity fail closed without allocating on the audio thread.

7. Reverb

A multi-comb / all-pass room reverb with pre-delay.

Parameter	Range	Default	Unit
Room size	0 … 100	20	—
Damping	0 … 100	30	—
Pre-delay	0 … 40	0	ms
Mix	0 … 50	10	%

Reverb mix above ~20 % can reduce call intelligibility.

8. Mix (global)

The final stage blends the preserved dry signal with the processed wet signal and applies output gain.

Parameter	Range	Default	Unit
Dry/Wet	0 … 100	50	%
Output gain	−12 … +12	0	dB

---

Preset catalog

The Preset Manager. Presets can be activated, renamed, duplicated, and imported/exported; per-app bindings are stored separately so presets stay portable.

Echidna ships with eight built-in presets. Badges: NAT natural, FX heavy effect, LL low-latency, HQ high-quality.

Preset	Badges	Character
Natural Mask	NAT, LL	+2 to +3 st, −100 to −200 formant cents, light comp, subtle EQ tilt — light disguise that stays natural.
Darth Vader	FX, LL	Pitch −6 to −8 st, formant −200 to −300 cents, low-pass ~3.5 kHz, short room reverb.
Helium	FX, LL	Pitch +5 to +7, formant +150 to +250, high-pass 160 Hz, +2 dB brightness @ 3 kHz.
Radio Comms	NAT, LL	Band-pass 300–3.4 kHz, mild 4:1 comp, tighter noise gate.
Studio Warm	HQ, NAT	Low-shelf +2 dB @ 120 Hz, high-shelf −1.5 dB @ 8 kHz, 2:1 comp, subtle plate reverb.
Robotizer	FX, HQ	Fast Auto-Tune (retune 10–20 ms, chromatic), fixed formant, wet ~80 %.
Cher-Tune	FX, HQ	Auto-Tune in a musical key, retune 1–5 ms, humanize 0–10 %, formant preserve on.
Anonymous	NAT, LL	Pitch −2, formant −150, de-ess EQ @ 6–8 kHz −3 dB, dry/wet 60 %.

Presets are stored as JSON (version: 1) with an engine block (latencyMode, blockMs) and a modules array keyed by effect id (gate, eq, comp, pitch, formant, autotune, reverb, mix). Per-app bindings are stored separately so presets stay portable. Presets can be created, renamed, duplicated, imported/exported (single or bundle), and shared.

---

Signed plugin system

Power users can extend libech_dsp.so with third-party effect modules. Plugins are loaded from the directory in the ECHIDNA_PLUGIN_DIR environment variable (default /data/local/tmp/echidna/plugins) and inserted into the chain immediately before the mix bus.

Each plugin ships two files:

1. <name>.so — a shared object exporting const echidna_plugin_module_t *echidna_get_plugin_module(), whose descriptor table declares one or more effects (identifier, optional display_name, version, flags, and create/destroy for an EffectProcessor instance). The module reports abi_version = ECHIDNA_DSP_PLUGIN_ABI_VERSION (currently 1).
2. <name>.so.sig — an Ed25519 signature over the raw .so payload.

The loader verifies the signature before dlopen, against a trusted public key baked in at build time (ECHIDNA_TRUSTED_PLUGIN_PUBKEY). The default is an all-zero fail-closed placeholder — a real key must be provided at build time to enable third-party plugins, and verification is only active when the DSP library is built with BoringSSL (ECHIDNA_HAS_BORINGSSL). A missing signature file, a failed check, or a build without a real key all cause the plugin to be rejected. Loaded plugins are prepared and reset whenever the engine reapplies a preset.

---

Safety & watchdog

The native bridge auto-bypasses a process when in-callback processing overruns for a sustained run of blocks, and a global Panic control instantly bypasses the engine for a hold window. Bypassed callbacks flag telemetry and increment XRuns for the Diagnostics view. Tuning is via ECHIDNA_WATCHDOG_US, ECHIDNA_WATCHDOG_CONSEC, ECHIDNA_BYPASS_MS, and ECHIDNA_PANIC_MS (see the developer guide). This keeps a heavy or misbehaving preset from destabilising a live call.