> 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. # Gemini 3.8 Flash-Lite TTS > Cost-efficient speech for voice agents and dubbing, with streaming, vocal tags, and two-speaker dialogue across 101 languages and 2,000+ voices. ![Gemini 3.8 Flash-Lite TTS](https://media.empiriolabs.ai/model-logos/gemini-3-8-flash-lite-tts.png) [Google](/providers/google) · Audio Generation `POST /v1/audio/speech` Cost-efficient speech for voice agents and dubbing, with streaming, vocal tags, and two-speaker dialogue across 101 languages and 2,000+ voices. ## At a glance | Field | Value | | ------------------- | ------------------------------------------------------------------------------------------------ | | Model id | `gemini-3-8-flash-lite-tts` | | Model release date | 2026-09-22 | | Input modalities | Text | | Output modalities | Audio | | Context window | - | | Weight precision | - | | Features | text_to_speech, multi_speaker, multilingual, voice_tags, voice_control, streaming, voice_cloning | | Native inference | No | | New | Yes | | Supported endpoints | `POST /v1/audio/speech`, `POST /v1/audio/speech:stream`, `GET /v1/voices` | | Alternate model ids | `gemini-3.8-flash-lite-tts`, `google/gemini-3.8-flash-lite-tts` | ## Pricing | Charge | Spec | Rate | | ------ | ----------------------- | ------- | | Input | per 1M prompt tokens | \$2.60 | | Output | per 1M generated tokens | \$31.20 | ## Example request ```bash curl https://api.empiriolabs.ai/v1/audio/speech \ -H 'Authorization: Bearer $EMPIRIOLABS_API_KEY' \ -H 'Content-Type: application/json' \ -d '{"model": "gemini-3-8-flash-lite-tts", "input": "Hello from EmpirioLabs."}' ``` ## Parameters | Parameter | Type | Required | Default | Description | | ------------------------- | ------ | -------- | ------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `input` | string | yes | - | Text to convert to speech. For multi-speaker mode, prefix lines with Speaker1: / Speaker2:. Up to 5,000 characters. Put delivery direction in style_prompt, and place these vocal tags inline where the sound should happen: \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \, \. Capitalize a word to stress it, and wrap a listener reaction in pipes, for example \|mm-hmm\|. · Max: 5000 | | `mode` | enum | no | `"single"` | single = one voice, multi = two-voice dialogue (uses voice + voice2 + speaker names). · Allowed: `single`, `multi` | | `language` | string | no | - | Optional BCP-47 language tag (en-US, es-ES, etc.). The language of the text is detected automatically. The 30 listed voices speak every supported language; a library voice must be used with its own language. | | `voice` | enum | no | `"Charon"` | Primary voice name (e.g. Kore, Puck, Aoede). Leave blank for the default. These 30 voices speak every supported language. GET /v1/voices lists the full library of more than 2,000 voices, and any voice id it returns is also accepted here. custom clones a voice from voice_audio_url and voice_consent_audio_url, and the response's voice_id then works here or in voice2 for 7 days. · Allowed: `Zephyr`, `Puck`, `Charon`, `Kore`, `Fenrir`, `Leda`, `Orus`, `Aoede`, `Callirrhoe`, `Autonoe`, `Enceladus`, `Iapetus`, `Umbriel`, `Algieba`, `Despina`, `Erinome`, `Algenib`, `Rasalgethi`, `Laomedeia`, `Achernar`, `Alnilam`, `Schedar`, `Gacrux`, `Pulcherrima`, `Achird`, `Zubenelgenubi`, `Vindemiatrix`, `Sadachbia`, `Sadaltager`, `Sulafat`, `custom` | | `voice_audio_url` | string | no | - | voice=custom only: an https URL to a 10 to 30 second recording of clean, natural speech from the voice to clone (WAV, MP3, M4A, OGG, WEBM or FLAC). Record it in a quiet room, on the same microphone as the consent recording. | | `voice_consent_audio_url` | string | no | - | voice=custom only: an https URL to the same speaker reading this statement aloud: "I am the owner of this voice and I consent to Google using this voice to create a synthetic voice model." The same statement in any of 30 languages is accepted; consent_statements in this parameter's metadata lists each one. | | `voice2` | enum | no | `"Kore"` | Second voice name for multi-speaker mode. · Allowed: `Zephyr`, `Puck`, `Charon`, `Kore`, `Fenrir`, `Leda`, `Orus`, `Aoede`, `Callirrhoe`, `Autonoe`, `Enceladus`, `Iapetus`, `Umbriel`, `Algieba`, `Despina`, `Erinome`, `Algenib`, `Rasalgethi`, `Laomedeia`, `Achernar`, `Alnilam`, `Schedar`, `Gacrux`, `Pulcherrima`, `Achird`, `Zubenelgenubi`, `Vindemiatrix`, `Sadachbia`, `Sadaltager`, `Sulafat` | | `speaker1_name` | string | no | `"Speaker1"` | Display name used in the input prefix for speaker 1 (default: Speaker1). | | `speaker2_name` | string | no | `"Speaker2"` | Display name used in the input prefix for speaker 2 (default: Speaker2). | | `output_format` | enum | no | `"WAV"` | Audio file format: WAV, MP3, OGG (Opus), or ALAW / MULAW for telephony. · Allowed: `WAV`, `MP3`, `OGG`, `ALAW`, `MULAW` | | `speed` | number | no | `1.0` | Playback rate. 1.0 = natural; \<1 slower, >1 faster. · Range: 0.25 – 2.0 | | `volume_gain` | number | no | `0` | Output gain in dB. 0 = unchanged. · Range: -96 – 16 | | `sample_rate` | enum | no | `"24000"` | Output sample rate in Hz (8000, 16000, 22050, 24000, 44100, or 48000). · Allowed: `8000`, `16000`, `22050`, `24000`, `44100`, `48000` | | `style_prompt` | string | no | - | Natural-language style direction (e.g. "warm, conversational" or "newscaster, serious"). | ## Notes **Limits** * Text: up to 5,000 characters per request * Audio billing: 32 output tokens per second of generated audio * 101 languages. The language of the text is detected automatically. **Voices** * The 30 voices in the `voice` list speak every supported language. * `GET /v1/voices?model=gemini-3-8-flash-lite-tts` lists the full library of more than 2,000 voices across 30 locales, with the language, accent, gender, pitch, and a description of each. Filter it with `language`, `gender`, `pitch`, or `q`. Any id it returns works as `voice` or `voice2`; when you also set `language`, use the voice's own language. **Custom voices** * Set `voice` to `custom` and pass two recordings of the same adult speaker as https URLs: `voice_audio_url`, 10 to 30 seconds of natural speech, and `voice_consent_audio_url`, the speaker reading "I am the owner of this voice and I consent to Google using this voice to create a synthetic voice model." The statement is accepted in 30 languages; `consent_statements` on that parameter lists each wording. * Record both in a quiet room on the same microphone. WAV, MP3, M4A, OGG, WEBM and FLAC are accepted, up to 15 MB each. * The response includes `voice_id` and `voice_id_expires_at`. Pass the `voice_id` as `voice` or `voice2` to reuse the voice for 7 days without new recordings. Anyone with the id can use the voice until it expires, so keep it private. **Delivery direction** * Put direction for the whole passage in `style_prompt` rather than in the text, for example "warm, unhurried, and reassuring". * Place vocal tags inline where the sound should happen. Use these English tags even when the text is in another language: ```text ``` * Capitalize a word to stress it. Wrap a listener reaction in pipes, for example |mm-hmm|, to add it without starting a new turn. **Multi-speaker** * Up to two speakers per request. Start each line with a speaker name and a colon, for example Speaker1: and Speaker2:. The names must match `speaker1_name` and `speaker2_name`. **Output** * WAV, MP3, OGG (Opus), ALAW, or MULAW, at 8,000 to 48,000 Hz. * `POST /v1/audio/speech:stream` sends 16-bit PCM chunks at 24,000 Hz while the audio is generated, then the complete file in the requested format. `speed` is available on `POST /v1/audio/speech` only. --- *Machine-readable schema:* `GET https://api.empiriolabs.ai/v1/models/gemini-3-8-flash-lite-tts`.