Text-to-Speech
Generate audio from text using multiple models and voices.
The TTS endpoint converts text into spoken audio. You send a string, choose a model and output format, and receive an audio file. Long inputs are automatically split into chunks and concatenated so you do not need to manage text length yourself.
Endpoint
POST /v1/audio/speechSee the Gateway API Reference for the full request and response schema.
Models
| Model | Provider | When to use |
|---|---|---|
qwen3-tts-flash | Qwen / DashScope | General-purpose TTS. Fast, good quality, no reference audio needed. |
qwen3-tts-vc-2026-01-22 | Qwen / DashScope | Voice clone variant. Requires a voice_clone_profile_id. Use when you need output that matches a custom voice. |
Output formats
| Format | Notes |
|---|---|
mp3 | Compressed, widely supported. Good default for web and mobile. |
opus | Low-latency compressed format. Good for real-time playback over networks. |
aac | Compressed, native on Apple platforms. |
flac | Lossless compression. Use when quality matters more than file size. |
wav | Uncompressed PCM in a WAV container. Largest files, no quality loss. |
pcm | Raw PCM samples with no header. Use when you are feeding audio into another processing pipeline. |
Parameters
Beyond the required input (text) and model, TTS accepts several optional parameters that control voice characteristics:
| Parameter | Type | Description |
|---|---|---|
speed | float | Playback speed multiplier. Values below 1.0 slow down, above 1.0 speed up. |
temperature | float | Controls randomness in speech generation. Lower values produce more consistent output. |
topP | float | Nucleus sampling threshold. Restricts token selection to the most probable candidates. |
stability | float | How consistent the voice stays across the generated audio. Higher values reduce variation. |
similarity | float | How closely the output should match the target voice. More relevant with voice cloning. |
style | float | Expressiveness of the delivery. Higher values produce more animated speech. |
enhancement | bool | Applies post-processing to improve audio clarity. |
speakerBoost | bool | Amplifies the speaker's presence in the mix. |
Not every parameter is supported by every model. When an unsupported parameter is passed, it is ignored rather than causing an error.
Text chunking
TTS models have a per-request text limit of approximately 500 characters. When your input exceeds this limit, Everstack automatically:
- Splits the text into chunks at sentence boundaries (or at the 500-character limit if no boundary is found)
- Sends each chunk to the provider as a separate synthesis request
- Concatenates the resulting audio segments into a single output file
This happens transparently. You send the full text, and you receive one audio file back. The chunking logic preserves natural sentence breaks so the stitched output sounds continuous.
When to use TTS
- Notifications and alerts that need an audio component alongside text
- Accessibility for visually impaired users or screen-reader alternatives
- Content production where you need to convert articles, documentation, or scripts into audio
- Agent responses where a voice interface is more natural than text
- Workflow automation where audio is an output artifact, such as podcast generation or voiceover creation
Using a cloned voice
To generate speech with a custom voice, set the voice_clone_profile_id field in your request and use the qwen3-tts-vc-2026-01-22 model. The profile must already be enrolled with the provider. See Voice Cloning for setup instructions.
Next steps
- Speech-to-Text covers the reverse direction, turning audio into text.
- Voice Cloning explains how to create and use custom voice profiles.
- Gateway API Reference has the full endpoint schema.