Audio generation

generateSpeech() turns text into audio bytes through a SpeechGenerationModel.

1. Generate speech

import { generateSpeech } from '@anvia/core/speech-generation'

const speech = await generateSpeech({
    text: 'Your incident summary is ready for review.',
    model: audioModel,
    voice: 'alloy',
    speed: 1,
    providerOptions: {
        response_format: 'mp3',
    },
    retries: {
        maxAttempts: 3,
    }
})

voice is required. speed defaults to 1 and must be a positive finite number. Provider voice names, formats, and speed ranges may be narrower than the core contract.

The response contains audio: { data: Uint8Array, mediaType?: string } and rawResponse. Read the bytes from speech.audio.data and the optional media type from speech.audio.mediaType.

2. Validate the final script

import { z } from 'zod'

const SpeechInput = z.object({
  text: z.string().min(1).max(3_000),
  voice: z.enum(['alloy', 'verse', 'aria']),
})

const input = SpeechInput.parse(requestedSpeech)

const speech = await generateSpeech({
    text: input.text,
    model: audioModel,
    voice: input.voice,
    speed: 1,
    providerOptions: {
        response_format: 'mp3',
    }
})

The voice allow-list is application policy, not a universal provider list. Do not send unbounded agent output directly into synthesis.

3. Store and serve the result

const asset = await mediaStore.put({
  bytes: speech.audio.data,
  mediaType: speech.audio.mediaType ?? 'audio/mpeg',
  metadata: {
    voice: input.voice,
    model: audioModel.modelId ?? 'unknown',
  },
})

Keep raw audio out of sessions, traces, and event logs. Use a worker for long scripts and batch generation. Add consent, disclosure, and review controls when synthetic audio represents a real person.

Next, process existing audio with transcription.