Gradium TTS
Gradium's public beta TTS model handles complex text natively—phone numbers, emails, IBANs, time expressions—with ultra-low latency and no preprocessing.
Overview
What is Gradium TTS
Gradium TTS is a high-fidelity text-to-speech model from Paris-based Gradium, released in public beta in July 2026. It natively synthesizes complex text like phone numbers, email addresses, and time expressions without requiring preprocessing, making it ideal for voice agents and real-time applications.
Running it privately on Venice
On Venice, Gradium TTS runs under an anonymized privacy tier—your text prompts are not stored or profiled. This means developers can use a powerful, low-latency TTS model without sacrificing user privacy or creating a data retention risk. You get the same production-ready quality, used privately and at predictable per-character pricing.
Assessment
Strengths and limitations
- Handles complex text natively—phone numbers, emails, IBANs, URLs, and time expressions—without preprocessing or normalization.
- Ultra-low latency with time-to-first-audio (TTFA) around 155–258 ms, ideal for real-time voice agents and conversational interfaces.
- Supports streaming via WebSocket for real-time synthesis, enabling low-latency delivery in agent loops and live UIs.
- Offers high-quality output in multiple formats including PCM, WAV, Opus, and telephony-optimized codecs.
- Compatible with Gradium’s voice library and custom voices, allowing for brand-consistent voice experiences.
- Not open-source or open-weights—cannot be self-hosted or fine-tuned by users.
- Limited language support compared to global giants—currently focused on English, French, Spanish, German, and Portuguese.
- Still in public beta, so edge cases in pronunciation or formatting may require feedback and iteration.
- No end-to-end encryption or TEE execution on Venice, limiting its use in high-assurance security environments.
Samples
Sample outputs
Generated on Venice with our standard prompt suite — the same scripts we run through every model of this type, so you can judge it like-for-like.
“On Venice, your prompts are processed privately and never stored, profiled, or used to train anyone's model.”
“Wait — so I can run a private voice model with zero data retention, and pay only for what I use? That's genuinely useful.”
“Three… two… one… liftoff! The rocket roared into the night sky as the crowd erupted in cheers.”
Specifications
Datasheet
- Maker
- Gradium
- Released
- July 2026
- Modality
- Text-to-Speech
- Open weights
- No
- Privacy on Venice
- Anonymized — prompts not stored
- Available on Venice since
- Jun 2026
- License
- Proprietary
API
Call it from your code
Venice exposes an OpenAI-compatible API. Point your base URL at Venice and pass the model id.
curl https://api.venice.ai/api/v1/audio/speech \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "tts-gradium-v1",
"input": "On Venice, your prompts are processed privately.",
"voice": "af_sky",
"response_format": "mp3"
}' --output speech.mp3Pricing
What it costs on Venice
Billed per character on Venice: $47.50 per 1M characters of synthesized speech.
New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.
Alternatives
How it compares
| Model | Best for | Price per 1M chars | Open weights | Streaming support | Voice cloning |
|---|---|---|---|---|---|
| Gradium TTS | Balances cost, quality, and low-latency streaming with native handling of complex text. | $47.50 / 1M chars | No | Yes | Yes |
| Chatterbox HD (Resemble AI) | Slightly more expensive, strong in voice cloning and expressive speech. | $50 / 1M chars | No | Yes | Yes |
| ElevenLabs Turbo v2.5 | Higher cost but known for emotional expressiveness and voice quality. | $62.50 / 1M chars | No | Yes | Yes |
| Inworld TTS-1.5 Max | Much cheaper but may lack in naturalness and complex text handling. | $12.50 / 1M chars | No | Yes | Yes |
| Kokoro Text to Speech | Lowest cost and open weights, but limited in voice cloning and enterprise features. | $3.50 / 1M chars | Yes | Yes | No |
Balances cost, quality, and low-latency streaming with native handling of complex text.
Use cases
What it is good for
- 01Voice agents and virtual assistants that need to read out dynamic user data like phone numbers and reference codes.
- 02Customer service automation where accurate, natural pronunciation of complex inputs is critical.
- 03Real-time translation and speech-to-speech applications using Gradium’s full stack.
- 04Interactive voice response (IVR) systems requiring low-latency, high-clarity audio synthesis.
- 05Multilingual applications needing native-sounding time and number expressions across supported languages.
Prompting
Getting better results
Send raw text directly—no need to spell out phone numbers or email addresses; the model handles formatting natively.
Use the WebSocket API for streaming LLM output to minimize latency in agent conversations.
Specify `model_name: 'gradium-tts-beta'` in your setup message to access the latest improvements during the beta period.
Leverage custom voices and pronunciation dictionaries to maintain brand voice and handle domain-specific terms.
Version history
Public beta release with improved handling of complex text and streaming support.
FAQ
Frequently asked questions
Gradium TTS is a high-quality, low-latency text-to-speech model developed by Gradium, a Paris-based AI company spun out of Kyutai. Released in public beta in July 2026, it excels at synthesizing complex text like phone numbers, emails, and time expressions without preprocessing.
On Venice, Gradium TTS costs $47.50 per 1 million characters of synthesized speech. Pricing is billed per character, with no subscription or minimum commitment required.
No. Gradium TTS is a proprietary model and not open source. It is not free to use, though developers can access limited free credits during the public beta to test its capabilities.
Yes. Gradium TTS supports real-time streaming via WebSocket, making it ideal for low-latency applications like voice agents and live UIs where audio must be generated as text is produced.
Gradium TTS supports English, French, Spanish, German, and Portuguese. It handles native time expressions, number formats, and pronunciation rules in each language without requiring text normalization.
Yes. A key strength of Gradium TTS is its ability to natively pronounce complex inputs like international phone numbers, email addresses, IBANs, and URLs without requiring developers to pre-process or spell them out.
Gradium TTS offers lower latency and better handling of complex text at a lower price point, making it ideal for voice agents. ElevenLabs Turbo v2.5 excels in emotional expressiveness and voice quality, better suited for content creation.
On Venice, Gradium TTS runs under an anonymized privacy tier—your prompts are not stored, profiled, or used for training. This allows you to use the model in compliance with privacy standards without sacrificing performance.
Run Gradium TTS privately
No prompt logging. No data used for training.