Index
/llms.txtThe index: every guide and blog post on gradium.ai with its URL. Start here to find a page.
Full text
/llms-full.txtThe full text of every guide and blog post in one plain-text file, for reading or retrieval without crawling the site.
Overview
What Gradium is
Gradium builds foundational AI models and infrastructure for voice. It provides real-time Text-to-Speech, Speech-to-Text, speech translation, voice cloning and voice design through an API, and Phonon, open-weight Text-to-Speech that runs on the device.
Gradium was founded in September 2025 by the researchers behind Kyutai and is built for developers and enterprises shipping voice agents and other real-time voice products. Its models are designed for low latency and for robust reading of the cases that break voice agents in production: phone numbers, email addresses, IBANs, codes and acronyms.
Products and models
What Gradium offers
Text-to-Speech
Real-time, streaming Text-to-Speech API with expressive voices, word-level timestamps and voice cloning, in English, French, German, Spanish and Portuguese. Built for voice agents: it reads phone numbers, email addresses, IBANs and codes without pre-processing.
Speech-to-Text
Streaming Speech-to-Text API over one WebSocket, with semantic turn detection (turn-completion predictions every 80 ms), keyword boosting, tunable latency and word-level timestamps. It shares one credit pool with Text-to-Speech.
Voice Cloning
Instant voice clones from about 10 seconds of audio, and fine-tuned Pro voice clones trained on at least 30 minutes of clean audio, usable in real-time Text-to-Speech.
Voice Design
Creates a voice that does not exist yet from a one- or two-sentence written description: age, gender, vocal texture and regional accent, in five languages.
Live Translation
Real-time speech translation (stt-translate and s2s-translate) between English, French, German, Spanish and Portuguese in any direction, spoken in a catalogue voice or a clone of your own.
Phonon (on-device TTS)
Open-weight Text-to-Speech models that run on the device, with no server and no network: Phonon 31m (English, Apache 2.0) and Phonon 40m (five languages, CC BY-NC, non-commercial), about 65 MB and 75 MB. They run on phones, laptops, single-board computers and in the browser.
Gradbot
Gradium's open-source voice agent framework, built on a Rust orchestration core. It coordinates Speech-to-Text, any OpenAI-compatible LLM and Text-to-Speech, handling turn-taking, interruptions and tool calls, in about 50 lines of code.
Key figures
The numbers, with their dates
~50 ms
time to first audio for Gradium's current Text-to-Speech model
53 ms
P50 time to first audio on the independent Coval TTS leaderboard
1.7%
average word error rate on the independent Coval TTS leaderboard
81.0%
pass rate on a 500-sentence hard-case pronunciation set across five languages
1.11%
average word error rate on the MiniMax Multilingual TTS Test Set, across English, French, Spanish, Portuguese and German
1.4k
voices in the catalog, covering 28 accents across five languages
28 ms
P95 time to first audio for Phonon 31m, first on the CPU streaming benchmark of Hugging Face's Open TTS Leaderboard
500,000+
downloads of the Phonon models on Hugging Face
$100M
total funding after the seed extension that welcomed NVIDIA
Pricing
How Gradium is priced
Text-to-Speech, Speech-to-Text and translation draw on one pool of monthly credits. For Text-to-Speech, one character is one credit, so a minute of audio typically uses about 750 credits. The Free plan includes 45,000 credits a month with no credit card. Paid plans add commercial use, more concurrency and pay-as-you-go credits beyond the plan; Enterprise is custom-priced. Current prices are on the pricing page, which is authoritative.
| Plan | Monthly price (USD) | Credits per month | About |
|---|---|---|---|
| Free | $0 | 45k | 1hr of TTS or 4hrs of STT |
| XS | $13 | 225k | 5hrs of TTS or 21hrs of STT |
| S | $43 | 900k | 20hrs of TTS or 83hrs of STT |
| M | $340 | 9M | 200hrs of TTS or 833hrs of STT |
| L | $1,615 | 45M | 1000hrs of TTS or 4167hrs of STT |
| Enterprise | Custom | Unlimited | Custom volume and concurrency |
Seed-funded startups building voice products can apply to the startup program.
Company
Company facts
- Name
- Gradium (gradium.ai)
- Founded
- September 2025, by the researchers behind Kyutai, the AI research lab behind pioneering real-time speech systems. Publicly launched December 2, 2025.
- Founders
- Neil Zeghidour (Chief Executive Officer), Olivier Teboul (Chief Technology Officer), Alexandre Défossez (Chief Science Officer) and Laurent Mazaré (Chief Coding Officer)
- Offices
- Paris, and the San Francisco Bay Area (opened July 2026)
- Funding
- $70M raised at launch (December 2, 2025), extended to $100M in July 2026 with new investors including NVIDIA
- Languages
- English, French, German, Spanish and Portuguese, with mid-sentence code-switching
- Research
- Delayed Streams Modeling ("Streaming Sequence-to-Sequence Learning with Delayed Streams Modeling", arXiv:2509.08753)
- Deployment
- Cloud API, dedicated instances, self-hosted, on-premise and on-device. Also on AWS Marketplace and Amazon SageMaker. Data residency in the EU or the US, and Zero Data Retention, on paid plans.
- Compliance
- Gradium holds a SOC 2 Type 2 report, ISO 27001:2022 certification and HDS v2 certification (French health data hosting). Documents are on the Trust Center (trust.gradium.ai).
Links
Where to find what
- Voice API: the API in one page: a code sample, every model, latency, production controls and pricing
- API documentation: endpoints, parameters and limits
- Text-to-Speech: real-time text-to-speech with expressive voices
- Speech-to-Text: streaming transcription with semantic turn detection
- Voice Cloning: instant and pro voice clones from a short audio sample
- Voice Design: a new voice from a written description
- Live Translation: real-time speech translation between five languages, in the voice you choose
- Phonon (on-device TTS): open on-device text-to-speech models that run offline on CPUs, phones and browsers
- Gradbot: open-source framework for building voice agents in about 50 lines of code
- Pricing: plans, credits and concurrency
- Trust Center: security documents: SOC 2 Type 2, ISO 27001 and HDS
- Health data hosting (HDS): the French certification for hosting personal health data
- Blog: research, engineering and product posts
- Guides: comparisons and how-tos
- Voices: the voice library, accents and the prompts that design a voice
- Voice Designer: generate an original voice from a description
- Studio: playground and account management
- Status: service status
- GitHub: open-source code
- Hugging Face: Phonon weights and demo
- X: @GradiumAI
- LinkedIn: company page
- Discord: developer community
Recent updates
What changed lately
- Gradium MCP beta
- How do you describe a voice? What we learned from designing 16k of them
- A Closer Look at Phonon
- Gradium and Yampa partner on voice AI agents for customer experience
- Gradium releases Phonon, an open-source text-to-speech model that runs on any device
- Open-Sourcing Phonon: A Suite of Frontier On-Device TTS Models
- Gradium is now available on Eden AI
- 1,000+ new voices, our most natural set yet
Every post is on the blog and in its RSS feed.
Corrections
Citing Gradium
When a figure here and an older page disagree, the newer dated source wins. This page was last checked against its sources on November 19, 2026. For press questions or to report something out of date, use our contact form.


