Gemini 3.8 | Flash-Lite | Text to Speech.

  • Google
  • Text to Voice
  • $1.00 / 1M text input + $12.00 / 1M audio output

Gemini 3.8 Flash-Lite TTS turns text into expressive WAV speech with selectable voices and style control for narration, dialogue and read-aloud apps.

View API docs

What Gemini 3.8 | Flash-Lite | Text to Speech is good at.

  • Best for

    Where it fits

    Narration and video voiceovers: turn a prepared script into a WAV file for a product walkthrough, presentation or short explainer. For example, text can…

  • Limits and tips

    Getting good results

    Write the transcript as the listener should hear it, with sentence punctuation and clear speaker turns. Spell out ambiguous abbreviations when a particular…

  • Related model

    Audio Mastering

    Run it on the same input to compare output, speed and price.

    View model

Model documentation

About this model

Gemini 3.8 Flash-Lite TTS turns text into expressive WAV speech with selectable voices and style control for narration, dialogue and read-aloud apps.

Overview

Gemini 3.8 | Flash-Lite | Text to Speech Overview

Gemini 3.8 Flash-Lite TTS generates spoken audio from a written script on each::labs. Developed by Google, it supports narration, read-aloud applications and short scripted dialogue.

The integration returns a complete WAV file with 16-bit PCM audio at 24 kHz in mono. Choose from 30 prebuilt voices and 24 listed locale codes. Keep the words to be spoken in text and put delivery instructions in the optional prompt field.

Single-speaker mode reads the script with one selected voice. Multi-speaker mode assigns two configured voices to named dialogue turns. A style prompt can guide delivery, but it is not a numeric pitch, speed or volume control.

Flash-Lite is the cheaper option: its audio-output rate is $12.00 per million tokens, compared with $18.00 for Flash. The text-input rate is $1.00 per million tokens for both.

Capabilities

Capabilities

  • Generate a complete 24 kHz, mono, 16-bit PCM WAV file from a written transcript.
  • Select one of 30 prebuilt voices; Kore is the single-speaker default.
  • Choose from the 24 locale codes listed in the integration schema; en-US is the default.
  • Guide vocal delivery with an optional natural-language style prompt.
  • Generate single-speaker narration or dialogue with up to two configured speakers.
  • Assign dialogue turns through speaker-alias prefixes that match the configured voices.

Use cases

Use Cases for Gemini 3.8 | Flash-Lite | Text to Speech

Narration and video voiceovers: turn a prepared script into a WAV file for a product walkthrough, presentation or short explainer. For example, text can contain “Welcome to the setup guide,” while prompt supplies “Friendly and clear.” Add the returned file to your video in a separate editing step.

Read-aloud features: give written instructions or short articles an audio version. Select a listed language code and review the result for proper names and technical terms. The request synthesizes the supplied transcript; it is not a translation task.

Scripted dialogue: produce a two-person introduction, training exchange or character conversation with two named speakers. Set the aliases and voices explicitly, then put each labelled turn on its own line.

Learning materials: create short lesson narration and spoken examples for review or later assembly. Keep requests within the byte limits. This model does not add subtitles, synchronize lips, mix music or provide a live conversation session.

Tips & tricks

Tips and Tricks

Write the transcript as the listener should hear it, with sentence punctuation and clear speaker turns. Spell out ambiguous abbreviations when a particular reading matters. Keep the first draft short enough to check pronunciation, pacing and voice choice before preparing a longer sequence.

Use prompt for a concise natural-language direction. For narration, try “Calm and clear, with a friendly tone.” For an instructional clip, try “Read clearly, with a brief pause between steps.” These are delivery suggestions, not guarantees of an exact speaking speed.

For dialogue, set mode to multi and configure distinct speaker aliases. A transcript such as “Speaker1: Welcome. Speaker2: Thank you.” must use aliases matching the request fields. voice_name selects the first voice; speaker2_voice selects the second.

Measure UTF-8 byte lengths before submission, especially for non-Latin scripts. Split longer material at sentence or scene boundaries and combine completed files in a separate editing workflow. Do not add SSML or unsupported numeric voice controls.

Technical spec

Technical Specifications

  • Input: required, non-empty text string and an optional style prompt.
  • Request limits: text up to 4,000 UTF-8 bytes; prompt up to 4,000 UTF-8 bytes; combined text and prompt up to 8,000 bytes. Count bytes, not characters.
  • Output: complete WAV file, 16-bit PCM, 24 kHz, mono; no MP3 selection or live audio streaming in this model integration.
  • Modes: single (default) or multi, with up to two speakers.
  • Voices: Achernar, Achird, Algenib, Algieba, Alnilam, Aoede, Autonoe, Callirrhoe, Charon, Despina, Enceladus, Erinome, Fenrir, Gacrux, Iapetus, Kore, Laomedeia, Leda, Orus, Pulcherrima, Puck, Rasalgethi, Sadachbia, Sadaltager, Schedar, Sulafat, Umbriel, Vindemiatrix, Zephyr, Zubenelgenubi.
  • Listed locales: ar-EG, bn-BD, nl-NL, en-IN, en-US, fr-FR, de-DE, hi-IN, id-ID, it-IT, ja-JP, ko-KR, mr-IN, pl-PL, pt-BR, ro-RO, ru-RU, es-ES, ta-IN, te-IN, th-TH, tr-TR, uk-UA, vi-VN.
  • Defaults: voice_name=Kore, language_code=en-US, speaker1_alias=Speaker1, speaker2_alias=Speaker2, speaker2_voice=Charon.
  • Temperature: not an accepted input parameter.
  • Pricing: $1.00 / 1M text input tokens + $12.00 / 1M audio output tokens.

Things to be aware of

Things to Be Aware Of

Count UTF-8 bytes rather than characters when preparing text and style instructions. Keep each within 4,000 bytes and their combined size within 8,000 bytes. Split larger scripts into separate requests instead of relying on truncation.

In multi-speaker mode, ensure each turn label matches the configured alias. A mismatched label can assign speech incorrectly. Put delivery direction in prompt, not in the spoken transcript. The finished audio should be checked for pronunciation and intended speaker assignment before publication.

Do not send temperature: this model's input schema does not accept it. Features documented for other Google speech APIs are not automatically available in this integration.

Key considerations

Key Considerations

Choose the voice and language code before preparing the full script. Locale support here means the 24 options exposed by this integration, not every language or feature available elsewhere in Google's speech products. Voice names must match the listed options exactly.

Keep the transcript separate from the delivery instruction. Use prompt for a direction such as “Calm and clear”; do not place that instruction in text unless you want it spoken. This integration creates a finished file, so design playback around the completed result rather than streaming chunks.

Flash-Lite has a lower audio-output price than Flash; it does not change the format or expand the exposed controls.

There is no promised generation latency or exact clip duration. Review the generated speech before including it in a published asset.

Limitations

Limitations

This integration does not expose SSML, live streaming, MP3 output, voice cloning, custom voice creation, or numeric pitch, speaking-rate and volume controls. It returns 24 kHz mono WAV, not 48 kHz or stereo.

Use only the 30 listed voices and 24 listed locale codes. Dialogue supports at most two speakers. Text and prompt each have a 4,000-byte request limit, with an 8,000-byte combined limit. Style instructions are generative directions and do not guarantee exact timing or identical delivery across requests.

Questions about this model

The short answers. For anything specific to your setup, talk to our team.

Talk to an engineerPlan for your expected volume.
  • Gemini 3.8 Flash-Lite TTS converts a written transcript into synthesized speech with selectable voices and style direction. It returns WAV audio for narration, read-aloud features and dialogue. The integration on each::labs accepts text, an optional style prompt and voice settings while keeping audio delivery consistent across requests.

Ready to close the gap?

Find the model. Build the workflow. Ship the result.