#!/usr/bin/env bun
/**
* Generate the labeled Voice Workbench corpus (#8785) to disk: for every
* built-in scenario, write `audio.wav` (the synthesized + acoustically degraded
* stream) and `ground-truth.json` (per-turn labels: speaker, transcript, respond
* decision, entity, applied environment) under a versioned corpus directory.
*
* bun run scripts/generate-voice-corpus.ts [--out
] [--meeting-stress]
*
* The synthetic (formant) path needs no models — it produces a reproducible,
* audible corpus a reviewer can LISTEN to (you can hear the noise/reverb/
* far-field/low-quality degradation on the robustness scenarios) and that the
* real-model lane scores WER/DER against. A real-TTS corpus is produced by
* injecting a `CorpusTtsSynthesizer` (gated; not wired here).
*
* Outputs are generated artifacts — write them to a gitignored dir, don't commit.
*/
import { mkdirSync, writeFileSync } from "node:fs";
import path from "node:path";
import {
generateVoiceCorpus,
writeVoiceCorpus,
} from "../src/services/voice/corpus-generator.ts";
import {
buildMeetingAcousticStressMatrix,
type MeetingAcousticStressCase,
} from "../src/services/voice/meeting-acoustic-stress-matrix.ts";
import { VOICE_WORKBENCH_SCENARIOS } from "../src/services/voice/workbench-scenarios.ts";
interface CorpusManifestEntry {
scenarioId: string;
classes: string[];
durationSec: number;
turns: number;
degraded: boolean;
dir: string;
stress?: Omit;
}
async function main(): Promise {
const args = process.argv.slice(2);
const outIdx = args.indexOf("--out");
const outDir =
outIdx >= 0 && args[outIdx + 1]
? path.resolve(args[outIdx + 1])
: path.resolve("voice-corpus-output");
const meetingStress = args.includes("--meeting-stress");
mkdirSync(outDir, { recursive: true });
const stressMatrix = meetingStress
? buildMeetingAcousticStressMatrix()
: null;
const scenarios = stressMatrix
? stressMatrix.cases.map((entry) => entry.scenario)
: VOICE_WORKBENCH_SCENARIOS;
const stressByScenarioId = new Map(
(stressMatrix?.cases ?? []).map((entry) => [entry.scenario.id, entry]),
);
const manifest: CorpusManifestEntry[] = [];
for (const scenario of scenarios) {
const corpus = await generateVoiceCorpus(scenario);
const dir = path.join(outDir, scenario.id);
writeVoiceCorpus(corpus, dir);
const degraded = corpus.groundTruth.turns.some((t) => t.environment);
const stress = stressByScenarioId.get(scenario.id);
manifest.push({
scenarioId: scenario.id,
classes: scenario.classes,
durationSec: Number(corpus.groundTruth.durationSec.toFixed(3)),
turns: corpus.groundTruth.turns.length,
degraded,
dir: path.relative(outDir, dir),
...(stress
? {
stress: {
id: stress.id,
snrDb: stress.snrDb,
background: stress.background,
room: stress.room,
quality: stress.quality,
speechStructure: stress.speechStructure,
speakerCount: stress.speakerCount,
expectedBehavior: stress.expectedBehavior,
seed: stress.seed,
sourceManifestIds: stress.sourceManifestIds,
},
}
: {}),
});
process.stdout.write(
` ${scenario.id.padEnd(28)} ${corpus.groundTruth.durationSec.toFixed(2)}s ${corpus.groundTruth.turns.length} turns${degraded ? " [degraded]" : ""}\n`,
);
}
const manifestPath = path.join(outDir, "manifest.json");
writeFileSync(
manifestPath,
`${JSON.stringify(
{
schemaVersion: 1,
mode: meetingStress ? "meeting_stress" : "voice_workbench",
...(stressMatrix
? {
meetingAcousticStressMatrix: {
schemaVersion: stressMatrix.schemaVersion,
seed: stressMatrix.seed,
requirements: stressMatrix.requirements,
sourceManifests: stressMatrix.sourceManifests,
cases: stressMatrix.cases.map(({ scenario: _scenario, ...entry }) => entry),
},
}
: {}),
scenarios: manifest,
},
null,
2,
)}\n`,
);
process.stdout.write(
`\n[corpus] wrote ${manifest.length} scenarios to ${outDir}\n[corpus] manifest: ${manifestPath}\n`,
);
}
main().catch((err: unknown) => {
process.stderr.write(`${err instanceof Error ? err.stack : String(err)}\n`);
process.exit(1);
});