@seanhogg/builderforce-voice

Voice-cloning + LLM-narration client for Builderforce.ai Studio. One seam — resolveNarrationEngine() (or the VoiceClient wrapper) — turns a voiceId (a studio_voice_clones.id) plus text into audio, choosing the best available backend and degrading honestly:

1. On-device clone (clone-client) — the @seanhogg/builderforce-studio SSM clone engine on WebGPU. $0 marginal infra, runs on the user's GPU, nothing is billed. Preferred when available.
2. Server clone (clone-server) — POST /api/studio/voice-clones/:id/synthesize on the Builderforce gateway: license-checked, token-metered (voice_clone_synthesis), read-through cached, persisted to studio_voiceovers. Works on every device, including mobile / Safari with no WebGPU.
3. Named fallback (fallback) — your existing non-cloned voice (Kokoro / Piper / Web Speech). Used only when no clone path can run, and always flagged cloned: false with a human-readable reason, so the UI can say "Cloning unavailable — using Narrator" instead of swapping silently.

Every studio LLM flow (AI script → narration, dubbing, the value-prop / pitch builder) routes through the same seam, so none of them re-implements provider selection, licensing, or the fallback contract.

Install

npm install @seanhogg/builderforce-voice
# optional, only for the on-device ($0) path:
npm install @seanhogg/builderforce-studio

Quick start (server path — works everywhere)

import { VoiceClient } from '@seanhogg/builderforce-voice';

const voice = new VoiceClient({ apiKey: process.env.BUILDERFORCE_API_KEY! });

const result = await voice.narrate('clone_abc123', {
  text: 'The AI wrote this — and it speaks in my voice.',
});
// result.engineId === 'clone-server', result.cloned === true
// result.audioUrl, result.durationMs, result.wordTimestamps

On-device clone (free, when WebGPU is present)

import { VoiceClient } from '@seanhogg/builderforce-voice';
import { VoiceCloneEngine } from '@seanhogg/builderforce-studio';

const engine = new VoiceCloneEngine();
const speaker = engine.enroll(referencePcm); // enrol once, persist the embedding

const voice = new VoiceClient({ apiKey, clientEngine: engine });
const result = await voice.narrate('clone_abc123', { text }, { speaker });
// Prefers 'clone-client' (on-device); never touches the metered endpoint.

The honesty contract

import { resolveNarrationEngine, getEngineUnavailableReason } from '@seanhogg/builderforce-voice';

const engine = await resolveNarrationEngine({ voiceId, providers, fallback });
if (!engine.cloned) showBanner(engine.fallbackReason); // shown BEFORE synthesizing
const audio = await engine.synthesize({ text });

getEngineUnavailableReason(providers) is the single source of truth for "can I clone right now" — the picker, the dubbing panel header, and the pitch button all read it instead of each recomputing availability.

Status

The server endpoint and the on-device clone model weights are tracked in the repo's Consolidated Gap Register (voice gaps #1991, #1994–#2000). This package is the client seam: it is complete and stable, and a better clone model is a config swap behind the same interface, not a call-site rewrite.