Start · SDK reference

One class, one session, one frame.

The JavaScript, Python and plugin surfaces are the same three objects with the same field names. If you learn @anecho/sdk, you have learned all of them.

new Anecho(options)

Constructor options
Option	Type	Default	Notes
apiKey	string	—	Optional. Used for hosted weight delivery and usage reporting; inference is on-device either way, so a local model runs without one.
engine	'local'	'local'	On-device inference. It is the only mode there is: no audio is sent anywhere.
modelBaseUrl	string	the manifest's URLs	Origin you serve weights from. Under Cross-Origin-Embedder-Policy a browser refuses a .onnx from a host that does not opt in, and your own origin does.
manifestUrl	string	—	Fetch the model manifest from here instead of the default.
manifest	Manifest	—	Pass a manifest object directly, for a build that ships its own.
cache	CacheMode	'auto'	Whether decoded weights are kept between sessions.
runtimeUrl	string	jsDelivr	onnxruntime-web UMD build for the inference worker. Self-host it in production — most CDNs do not promise the CORP header require-corp needs.
wasmPaths	string	jsDelivr	Directory holding ort-wasm*.wasm. Same caveat.
transport	Transport	best available	Force a transport instead of letting the SDK pick.
onProgress	(p) => void	—	Weight-loading progress, for a determinate progress bar.

anecho.createSession(options)

Session options
Option	Type	Default	Notes
model	string	'clearline-s'	Model id. See the table below.
sampleRate	number	8000	The rate of the blocks you will push. Nothing in the SDK resamples it — that is the point.
splitPipeline	boolean	true	Emit enhanced and raw as separate, sample-aligned outputs.
engine	string	—	Override the engine backend for this session.
vad	boolean	true	Emit per-block speech and score alongside the audio.
signal	AbortSignal	—	Abort model loading and tear the session down.

session.process(block)

signature

process(block: Float32Array): {
  enhanced: Float32Array;
  raw: Float32Array;
  speech: boolean;
  score: number;
  latencyMs: number;
}

Processed frame fields
Field	Type	Notes
enhanced	Float32Array	Processed audio, same length as the input block.
raw	Float32Array	The input block, delayed to match `enhanced` exactly. Never re-align by hand.
speech	boolean	Onset's decision for this block.
score	number	Nyquist's call-quality estimate, 0–1. Low values predict downstream failure.
latencyMs	number	Algorithmic latency of the configured model. Constant for a session.

Models

Model ids and what each one is for
Id	Model	Role	Variants
clearline	Clearline	FLAGSHIP — primary-speaker isolation at 8 kHz, enrollment-free, streaming	clearline-xsclearline-sclearline-m
chamber	Chamber	Wideband (16/48 kHz) speech enhancement + primary-speaker isolation	chamber-xschamber-schamber-m
onset	Onset	Voice activity detection and turn-taking	onset-xsonset-sonset-m
nyquist	Nyquist	Call-audio quality scoring and failure prediction	nyquist-xsnyquist-snyquist-m

Clearline is the flagship and the default. -xs is the browser-capable variant, -s the default for server-side real-time work, -m the quality-first variant for offline or generously-provisioned paths. Real-time factor and algorithmic latency for every engine we have measured — competitors included — are on the leaderboard. Our own rows land there the same way everyone else’s did, and not before.

Errors

Error codes
Code	HTTP	Meaning
invalid_key	401	The key is malformed, revoked, or scoped to a different environment.
rate_limited	429	Test keys are capped per minute. Retry with backoff; the audio path is unaffected.
unsupported_sample_rate	400	Pass 8000, 16000 or 48000 and match it to your transport.
model_not_found	404	Model id does not exist or is not enabled for this account.
quota_exhausted	402	Free-tier minutes are spent. Processing continues on the raw passthrough path rather than dropping your call.