7.4 billion tokens per month. 34 free LLM providers. 635 free model endpoints. All behind one /v1 endpoint, plus any custom OpenAI-compatible endpoint. Smart routing, automatic failover, encrypted keys. Personal experimentation only.
TypeScript
// Turns an analyser's frequency snapshot into bar heights for the dictation
// wave drawn inside the composer while the mic is live. Pure, so the shaping
// is tested without an AudioContext.
/** Bars across the field. Odd so one sits dead centre. */
export const WAVE_BAR_COUNT = 31
/**
* `data` is a getByteFrequencyData buffer (0..255 per bin). The bins are
* bucketed evenly into `bars`, each bucket's mean drives one bar, and the
* result is scaled into [min, max] px. Speech energy lives in the low bins,
* so the buckets are laid out from the centre outwards — the middle bars
* carry the voice, the edges the hiss — which reads as a wave rather than a
* left-heavy staircase.
*/
export function barHeights(data: ArrayLike<number>, bars: number = WAVE_BAR_COUNT, min = 3, max = 22): number[] {
const n = data.length
if (bars <= 0) return []
if (n === 0) return Array.from({ length: bars }, () => min)
// Only the lower half of the spectrum carries anything a microphone hears.
const usable = Math.max(1, Math.floor(n / 2))
const perBar = usable / bars
const levels: number[] = []
for (let b = 0; b < bars; b++) {
const start = Math.floor(b * perBar)
const end = Math.max(start + 1, Math.floor((b + 1) * perBar))
let sum = 0
for (let i = start; i < end && i < n; i++) sum += data[i]
levels.push(sum / (end - start) / 255)
}
// Centre-out ordering: bucket 0 (lowest, loudest) in the middle, then
// alternating right/left so energy falls off symmetrically.
const ordered: number[] = new Array(bars).fill(0)
const mid = Math.floor(bars / 2)
for (let b = 0; b < bars; b++) {
const step = Math.ceil(b / 2)
const idx = b % 2 === 0 ? mid + step : mid - step
ordered[idx] = levels[b]
}
return ordered.map(v => min + Math.min(1, Math.max(0, v)) * (max - min))
}