> For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://docs.empiriolabs.ai/llms.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.empiriolabs.ai/_mcp/server. # TTS 2 > Realtime voice model that takes plain-English delivery direction, holds one voice identity across 200+ languages, and streams with word timestamps. ![TTS 2](https://media.empiriolabs.ai/model-logos/tts-2.png) [Inworld](/providers/inworld) · Audio Generation `POST /v1/audio/speech` Realtime voice model that takes plain-English delivery direction, holds one voice identity across 200+ languages, and streams with word timestamps. ## At a glance | Field | Value | | ------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------- | | Model id | `tts-2` | | Model release date | 2026-05-05 | | Input modalities | Text | | Output modalities | Audio | | Context window | - | | Weight precision | - | | Features | multi\_speaker, real\_time, low\_latency, streaming, word\_timestamps, character\_timestamps, multilingual, expressive\_prosody, voice\_direction | | Native inference | No | | New | No | | Supported endpoints | `POST /v1/audio/speech`, `POST /v1/audio/speech:stream`, `GET /v1/voices` | | Alternate model ids | `inworld-tts-2` | ## Pricing | Charge | Spec | Rate | | --------- | ----------------- | --------------------- | | Synthesis | per 1M characters | \$22.00 (was \$25.00) | ## Example request ```bash curl https://api.empiriolabs.ai/v1/audio/speech \ -H 'Authorization: Bearer $EMPIRIOLABS_API_KEY' \ -H 'Content-Type: application/json' \ -d '{"model": "tts-2", "input": "Hello from EmpirioLabs."}' ``` ## Parameters | Parameter | Type | Required | Default | Description | | -------------------------- | ------ | -------- | --------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `input` | string | yes | - | Text to synthesize. Max 2,000 characters per request; chunk longer copy at sentence boundaries on the client. Optionally open the text with a bracketed delivery direction in plain English, for example `[whispering]` or `[Speak warmly, like you are greeting an old friend]`. The bracketed text steers delivery instead of being read aloud, and its characters count toward the billed character total. · Max: 2000 | | `voice` | enum | no | `"Sarah"` | Voice preset. 20 hand-picked voices covering English, Spanish, Portuguese, Hindi, and various accents. For the full voice catalog (including cloned voices), use voice\_id instead. · Allowed: `Sarah`, `Olivia`, `Elizabeth`, `Ashley`, `Wendy`, `Julia`, `Priya`, `Pixie`, `Deborah`, `Alex`, `Mark`, `Edward`, `Theodore`, `Ronald`, `Dennis`, `Timothy`, `Shaun`, `Craig`, `Hades`, `Heitor` | | `voice_id` | string | no | - | Free-form voice ID. Overrides voice when set. Use this to address any voice beyond the curated 20-preset list, including cloned and regional voices. Pass any voice name from GET /v1/voices. Example: Ashley, Mark. | | `language` | enum | no | `"en-US"` | BCP-47 language code. This model holds one voice identity across 200+ languages and locales, so you can pass any supported code here even if it is not in the dropdown. · Allowed: `en-US`, `en-GB`, `es-ES`, `es-MX`, `fr-FR`, `de-DE`, `it-IT`, `pt-BR`, `pt-PT`, `nl-NL`, `pl-PL`, `ru-RU`, `ja-JP`, `ko-KR`, `zh-CN`, `hi-IN`, `ar-EG`, `he-IL`, `tr-TR`, `vi-VN`, `th-TH`, `id-ID`, `uk-UA`, `el-GR`, `cs-CZ`, `sv-SE` | | `output_format` | enum | no | `"WAV"` | Audio container/codec. WAV = LINEAR16 inside RIFF (ubiquitous). MP3 / OGG = compressed. PCM = headerless raw, useful for chunked real-time playback. FLAC = lossless. · Allowed: `MP3`, `WAV`, `OGG`, `FLAC`, `PCM`, `ALAW`, `MULAW` | | `sample_rate` | enum | no | `"24000"` | Output sample rate in Hz. 24000 is Inworld's default and what their voice models train at; raise to 48000 for broadcast quality. · Allowed: `8000`, `16000`, `22050`, `24000`, `32000`, `44100`, `48000` | | `speed` | number | no | `1.0` | Speaking rate multiplier. 0.5 = half speed, 1.5 = 50% faster. · Range: 0.5 – 1.5 | | `temperature` | number | no | `1.0` | Voice expressiveness / variability. Lower = more consistent / "flat"; higher = more expressive but more variation between renders. · Range: 0.1 – 2.0 | | `bit_rate` | number | no | `128000` | Bitrate in bps for MP3 / OGG\_OPUS. Ignored for other encodings. · Range: 32000 – 320000 | | `apply_text_normalization` | enum | no | `"ON"` | When ON, Inworld expands numbers / abbreviations / dates into spoken form ("USD 5" → "five US dollars"). · Allowed: `ON`, `OFF` | | `timestamp_type` | enum | no | `"NONE"` | If non-NONE, the response includes per-word or per-character timestamps in timestamp\_info. Useful for caption / highlight UIs. · Allowed: `NONE`, `WORD`, `CHARACTER` | ## Notes **Limits** * Max input: 2,000 characters per request (chunk longer text at sentence boundaries) * WebSocket: 20 concurrent connections, 5 contexts/connection * Per-WS message: 1,000 characters **Voice direction** * Open the text with a bracketed instruction in plain English to steer delivery, for example `[whispering]`, `[excited]`, or `[Speak warmly, like you are greeting an old friend]` * The bracketed text is treated as direction and is not read aloud * Direction text counts toward the billed character total, so keep it short **Latency** * p90 TTFB: under 100 ms (Inworld benchmark) **Voices** * One voice identity held across 200+ languages and locales, including switching language mid-generation * Curated presets in the dropdown; pass any other voice ID via `voice_id` --- *Machine-readable schema:* `GET https://api.empiriolabs.ai/v1/models/tts-2`.