What this is

A Speechify TTS provider for the Vercel AI SDK — developers get Speechify speech generation with the same one-line interface as OpenAI / ElevenLabs / Deepgram:

import { speechify } from '@speechify/vercel';
import { generateSpeech } from 'ai';

const { audio } = await generateSpeech({
  model: speechify.speech('simba-english'),
  text: 'Hello!',
});

No Speechify provider exists on npm or the AI SDK community providers list today — the slot is open. ElevenLabs and Deepgram are already first-class there.

Architecture

The provider is a thin bridge over the official @speechify/api client — no hand-rolled HTTP or duplicated wire types. It only:

• maps AI SDK call options (text, voice, speed, outputFormat, providerOptions) onto Speechify.GetSpeechRequest,
• converts SpeechifyError → APICallError (user aborts re-thrown as the original AbortError so AI SDK retry/abort semantics hold),
• disables the SDK's internal retries (maxRetries: 0) because the AI SDK applies its own retry policy on top.

One deliberate extension: the live API accepts codec strings (pcm_16000, mp3_24000_128, ulaw_8000) via output_format, verified live, but @speechify/api@2.0.0's GetSpeechRequest doesn't have the field yet — the type is extended locally until the next SDK generation picks it up. Base URL defaults to the SDK's environment (api.speechify.ai).

Naming (per the Slack conventions)

• npm package: @speechify/vercel
• repo: Speechify-AI/vercel-sdk — transferred from SpeechifyInc/speechify-api-sdk-vercel on 2026-07-06 (GitHub redirects all old URLs).

State

Code-complete and tested. Implements the current SpeechModelV4 spec (ai@7) against POST /v1/audio/speech. TTS only — the API has no STT endpoint to wrap.

• ✅ Typecheck clean, 20/20 unit tests pass (incl. a guard that the bridge never double-retries), tsup build green (ESM+CJS+types), npm publish --dry-run clean
• ✅ Smoke test passed against the live API through the bridge (2026-07-06) with the defaults — model simba-3.2, voice geffen_32: mp3 generated, speech marks + billable count correct, bad-key path raises a clean AI_APICallError. (Earlier pre-bridge run also live-verified codec output_format pcm_16000, SSML passthrough, and speed→prosody wrap.) Defaults note: the simba-3.2 curated voice set was pulled live from /v1/voices (8 voices, *_32); geffen_32 chosen as the en-US voice with the broadest use-case tags.
• ✅ Release setup mirrors speechify-api-sdk-typescript: release-please on master, pnpm/corepack Node 24, OIDC provenance publish. One fix over the SDK repo: src/version.ts carries the x-release-please-version annotation so version stamping actually works

Design notes: speed is implemented via SSML <prosody> wrap (the API has no speed param); word-timing speech marks are exposed via providerMetadata.speechify. Full surface in the README.

Remaining steps (permission-gated — @luke-speechify)

1. Merge this PR (smoke test already re-run post-refactor with the final defaults, passing).
2. Transfer the repo — done 2026-07-06, this is now Speechify-AI/vercel-sdk.
3. First npm publish (must be manual once — trusted publishing can't be pre-configured for a package that doesn't exist): with @speechify scope rights, pnpm build && npm publish --access public.
4. Configure trusted publishing on npmjs.com → @speechify/vercel → Trusted Publisher: repo Speechify-AI/vercel-sdk, workflow release-please.yml. After that, releases are fully automated (merge release-please PR → CI publishes with provenance), same as @speechify/api. Note: repo needs to be flipped public for provenance — it's private pending your review.
5. Community listing: PR a page to vercel/ai under content/providers/05-community-providers/ (copy the structure of 41-soniox.mdx, another speech-only provider). Vercel explicitly welcomes these PRs — no approval gate.

Questions → Vrishab.