Inworld

Inworld covers the full speech surface — sync + streaming in both directions, both capability markers shipped — and adds router-style passthroughs that proxy to other STT providers (AssemblyAI, Soniox, Groq Whisper) under the same Inworld auth and billing.

Install

pnpm add @effect-uai/core @effect-uai/inworld effect

Layers

Layer	Registers	Capability markers
`@effect-uai/inworld/InworldTranscriber`	`InworldTranscriber` + `Transcriber`	— (sync)
`@effect-uai/inworld/InworldRealtimeTranscriber`	`InworldRealtimeTranscriber` + `Transcriber`	`SttStreaming`
`@effect-uai/inworld/InworldSynthesizer`	`InworldSynthesizer` + `SpeechSynthesizer`	— (sync + chunked NDJSON)
`@effect-uai/inworld/InworldRealtimeSynthesizer`	`InworldRealtimeSynthesizer` + `SpeechSynthesizer`	`TtsIncrementalText`

The sync and realtime layers are separate so you can pull only what you need — the realtime paths add WS / JWT plumbing.

import { Config, Effect, Layer } from "effect"
import { FetchHttpClient } from "effect/unstable/http"
import * as Socket from "effect/unstable/socket/Socket"
import { layer as realtimeTranscriber } from "@effect-uai/inworld/InworldRealtimeTranscriber"
import { layer as realtimeSynth } from "@effect-uai/inworld/InworldRealtimeSynthesizer"

const inworld = Layer.unwrap(
  Effect.gen(function* () {
    const apiKey = yield* Config.redacted("INWORLD_API_KEY")
    return Layer.mergeAll(realtimeTranscriber({ apiKey }), realtimeSynth({ apiKey }))
  }),
)

const mainLayer = inworld.pipe(
  Layer.provide(FetchHttpClient.layer),
  Layer.provide(Socket.layerWebSocketConstructorGlobal),
)

Models

STT

Model	Native	Streaming WS
`inworld/inworld-stt-1`	First-party (experimental)	✓
`assemblyai/universal-streaming-english`	AssemblyAI passthrough	✓
`assemblyai/universal-streaming-multilingual`	AssemblyAI	✓
`assemblyai/u3-rt-pro`	AssemblyAI	✓
`assemblyai/whisper-rt`	AssemblyAI	✓
`soniox/stt-rt-v4`	Soniox	✓
`groq/whisper-large-v3`	Groq	— (sync only)

Passthrough models are billed against your Inworld key — no separate contracts. Sync STT works for all; streaming WS is supported by everything except groq/whisper-large-v3.

TTS

Model	Latency (P50)	Languages	Notes
`inworld-tts-2`	~200 ms	100+	Flagship; honors `deliveryMode`
`inworld-tts-1.5-max`	~200 ms	15
`inworld-tts-1.5-mini`	~120 ms	15	Lowest latency

Voice IDs are human-readable names (“Sarah”, “Edward”, …) but Inworld doesn’t publish a list-voices REST endpoint — browse via the Inworld Portal. InworldVoiceId is typed as plain string.

Request shape

// TTS sync + streaming
type InworldSynthesizeRequest = {
  readonly model: InworldTtsModel
  readonly voiceId: InworldVoiceId
  readonly text: string // omitted on streamSynthesisFrom
  readonly outputFormat?: AudioFormat
  readonly speed?: number
  readonly temperature?: number // (0, 2]
  readonly deliveryMode?: "STABLE" | "BALANCED" | "CREATIVE" // tts-2 only
  readonly applyTextNormalization?: "ON" | "OFF" // default "ON"
}

deliveryMode is the style-steering knob on inworld-tts-2: STABLE for consistent reading-voice output, CREATIVE for more expressive prosody. Older models ignore it silently.

applyTextNormalization: "OFF" skips the server-side text rewriter (expanding numbers, abbreviations, etc.) — faster, but punctuation pacing is on you.

Wire / auth notes

Sync endpoints: REST with bearer auth via Authorization header. Sync TTS comes back as either a single base64 audio blob (/tts/v1/voice) or NDJSON one chunk per line (/tts/v1/voice:stream).
Realtime endpoints: short-lived JWT minted from the API key (wsAuth.ts), passed as a bearer on the WS upgrade. Realtime STT expects PCM s16le at 16 kHz mono.

Audio encoding options for TTS (audioConfig.audioEncoding): LINEAR16, MP3, OGG_OPUS, ALAW, MULAW, FLAC, PCM, WAV. Caveat: sync LINEAR16 / WAV responses include a WAV header; streaming chunks don’t. The codec layer surfaces this via AudioFormat.container ("wav" vs "raw").

Errors

Standard HTTP → AiError mapping. WS-side failures surface on the stream’s error channel; non-fatal mid-stream errors come through as TranscriptEvents with _tag: "error".