Push-to-Talk Voice Kit

Why this takes weeks to get right

Browser voice input works on the first try and then fails on real users' machines in a dozen ways that produce no errors:

• Key repeat fires ~30 times a second while a key is held; the naive handler restarts the recorder on every one and captures nothing.
• Blur while held means keyup never arrives — the mic stays open and the browser's recording indicator stays lit. Users notice that.
• The space bar is the natural push-to-talk key and also the key people press to type.
• getUserMedia is async, so a fast tap can release before the stream exists; drop that release and the recording never stops.
• Silence detection with one threshold cuts people off between words, because natural speech has genuinely silent gaps.
• Wrong sample rate gives you confident nonsense from the model rather than an error — the worst failure shape, because you blame the model.

What's in it

• `VoiceActivityDetector` — Schmitt-trigger thresholds, sustained-silence requirement, minimum turn length, adaptive noise floor, calibration pre-roll, and a stuck-turn safety valve.
• `PushToTalkMachine` — pure state machine covering repeat, blur, async start, fast taps, and permission denial. No DOM, so it's fully testable.
• `prepareForSpeechModel` — channel averaging, interpolated resampling to 16kHz, clamped PCM16, plus a WAV encoder.

Proof

22 tests, each named after the failure it prevents. Two of them exist because the suite caught real bugs in this package's own first draft: the noise floor adapted *before* detecting speech (so every turn ratcheted it upward until the detector went deaf to its own user), and once that was fixed a detector that mis-heard room tone could never recover — which is why calibration and the safety valve exist.

Delivery

Source delivered as a private repository invite within 24 hours of purchase. Single-product commercial license: use and modify in one shipped product; no redistribution or resale of the source.

Interface

What you call, and what comes back. Types and signatures only — the implementation ships with the source.

  export function toMono(channels: Float32Array[]): Float32Array;
  export function resampleLinear( input: Float32Array, fromRate: number, toRate: number, ): Float32Array;
  export function toPcm16(input: Float32Array): Int16Array;
  export function prepareForSpeechModel( channels: Float32Array[], sourceRate: number, targetRate = 16000, ): Int16Array;
  export function encodeWav(pcm: Int16Array, sampleRate = 16000): Uint8Array;
  export function rms(frame: Float32Array): number;

  export type PttState = "idle" | "starting" | "recording" | "stopping";
  export type VadState = "silent" | "speaking";