Tesserae.Lauschwerk 26.10.72899

Tesserae.Lauschwerk

On-device voice input for Tesserae applications: keep the user's voice and drop the rest, tell when someone speaks, hear any typed wake word, recognize a voice. Typed C# (Transpose) over the Lauschwerk web SDK - a Rust engine compiled to WebAssembly, which the package carries - plus ready-made Tesserae components. Everything runs in the browser - no server, no key, no telemetry.

Preview (0.x). Apache-2.0, code and weights. Where it comes from: the repository's README.

Install

<PackageReference Include="Tesserae.Lauschwerk" Version="..." />

The build copies the SDK into your site at assets/js/lauschwerk/ (about 620 KB, the engine included). Nothing is fetched until something needs it.

The models ship in the package (7.5 MB for all five, sha256-checked at build) and the build copies them to assets/js/lauschwerk/models/, where the SDK loads them from by default: nothing is downloaded from a third party, and no user sends anyone their IP. To serve them from somewhere else (a CDN):

LauschwerkJs.ModelBaseUrl = "https://models.example.com/lauschwerk";   // once, before the first input

A call (WebRTC)

using Tesserae.Lauschwerk;

var input = await LauschwerkJs.CreateCallInputAsync(new CallInputSettings { Learn = true, FarEnd = remoteTrack });
pc.addTrack(input.Track);             // only the user's voice goes out
input.Bind(pc.getSenders()[0]);
input.VoiceprintLearned += print => localStorage.setItem("print", print.ToJson());

A voice agent

var input = await LauschwerkJs.CreateAgentInputAsync(new AgentInputSettings { FarEndManual = true, Format = PcmFormat.Int16, SampleRate = 24000 });
input.Audio += pcm => socket.send(pcm);           // continuous 32 ms packets
input.Utterance += u => Transcribe(u.Pcm);        // one per sentence, 300 ms pre-roll
input.FarEndActive(true);                         // the agent starts speaking

A wake word

var word = await LauschwerkJs.CompileWordAsync("Hey Nova", SpeechLanguage.English);   // any word, 7 languages
var wake = await LauschwerkJs.CreateWakeListenerAsync(new WakeListenerSettings(word));
wake.Wake += e => StartListening(e.Audio);

Files

var r = await LauschwerkJs.ProcessAsync(file, new ProcessSettings { Focus = true, Cue = true });
r.Audio;   // AudioClip, 16 kHz, processed
r.Speech;  // where speech is
r.ToWav(); // Blob

Components

using static Tesserae.Lauschwerk.LauschwerkUI;

They draw with Tesserae's --tss-* tokens, so they follow light and dark mode and the custom themes.

Good to know

• A near microphone, one person. lauschwerk keeps the closest voice (or the voice of a print). 16 kHz voice: the output carries speech up to 8 kHz.
• Only a missing microphone fails an input (LauschwerkException.Code: mic.denied, mic.not-found, mic.busy, page.insecure, browser.unsupported). Models still loading, a failed download, a slow device: the input runs VoiceMode.Raw with a RawReason, and the SDK retries.
• Content Security Policy: script-src 'self' 'wasm-unsafe-eval'; worker-src 'self' - the worker and worklet are same-origin files, no blob: needed. Add the models' origin to connect-src if they are not on yours.
• Privacy: no network call besides the models. A voiceprint your app stores is biometric data (GDPR art. 9).

A sample gallery with a page per feature lives in the repository.