Gradium is an AI firm based in Paris that creates Audio Language Models for Voice AI and foundational voice AI technologies to enable voice interaction at scale using natural and expressiveness. Founded by Researchers from Google DeepMind, Meta and Jane Street in September 2025 Gradium was developed out of the French non-profit AI research laboratory Kyutai to develop and commercialize cutting edge voice AI research.
Kyutai, which is the parent organization, was established in June 2017 as the first major European non-profit AI research laboratory with backings from French billionaire investors and international investors to the tune of €300M.
Gradium was created in September 2025 by Kyutai researchers Neil Zeghidour, Olivier Teboul, Laurent Mazaré and Alexandre Défossez to commercialize their research into audio language models.
Gradium completed a $70 million Seed Round in December 2025 led by FirstMark Capital and Eurazeo with participation from DST Global Partners, Eric Schmidt, Xavier Niel and other prominent investors.
In December 2025, Gradium emerged from its 3 month Stealth Mode with production ready speech-to-text, text-to-speech, and voice cloning models and began to generate revenue just a few weeks after it was formed.
Cloud-based platform supporting both rapid prototyping and enterprise-scale deployments
Proprietary audio language models designed to deliver natural, expressive voice interactions with ultra-low latency; capable of speech-to-text, text-to-speech, voice cloning, and dialogue generation across multiple languages
Based on company press releases, product documentation, and public coverage. Specific cloud provider and infrastructure details not disclosed.
Gradium currently supports English, French, German, Spanish and Portuguese for both TTS and STT. More languages are being added to the platform. The same one voice works across all five languages with consistent pronunciation.
Gradium analyzes voice identity based on as few as 10 seconds of recorded audio, and generates your custom speech via the AI voice API or studio. On higher plans, you get up to 1,000 clones with high fidelity across languages.
Credit-based: A free tier with 45k credits (~1hr TTS/3hrs STT). Pay-as-you-go pricing starting at $4.0 on higher plans. Enterprise features include private cloud and no data retention. Commercial use begins on higher tiers.
Gradium emphasizes ultra-low latency real-time streaming and full-duplex conversations with semantic voice activity detection. In addition to ultra-low latency, it also supports instant cloning with improved speaker similarity per human evaluation and multilingual consistency.
Private Cloud and Zero Data Retention are included in enterprise plans and are intended to support compliance with privacy regulations in use cases such as health care. No certifications have been specifically identified.
Yes, it integrates with LiveKit, Pipecat and other major agent frameworks. Python and Rust client libraries are available. Deployment of SageMaker through the AWS Marketplace is supported.
Yes. Free tier (no credit card required) includes: 1. 45k credits; 2. Studio/API access; and 3. 5 instant voice clones for real-time testing of voice cloning.
Designed for systems that require latency, with stable performance at high concurrency levels; enables real-time streaming and full-duplex conversations which eliminate the traditional speaker turns of human conversation.
Gradium offers production-ready, ultra-low latency text-to-speech (TTS) & speech-to-text (STT) models with unique instant voice cloning capability and multi-language support positioning it as a serious contender for low-latency real-time voice AI applications. Funded by over $70 million dollars, founded by leading researchers in AI, and delivering expressive, scalable voice synthesis for agents and interactive experiences, its long-term enterprise reliability is still emerging due to being a recent stealth exit.
Gradium is ideal for developers and enterprises building low-latency, expressive real-time voice applications where multi-language cloning and streaming performance are critical.
Gradium provides multilingual TTS/STT in real time using a voice clone that can be made from an 10-second clip of voice, along with low-latency WebSocket APIs, and will integrate into agent platforms. Received $70 million in seed funding after going “stealth”, was created by some of the world’s top AI researchers working at Kyutai/DeepMind, and uses their audio language model (ALM) technology which has been proven to perform better than traditional TTS pipeline architectures. Offers free tier for rapid prototyping and supports use cases such as gaming, healthcare, and customer service.
Good - detailed from official site, funding announcements, and tech reviews. Limited API docs depth, pricing beyond tiers requires contact, no public status page.
Launched from stealth December 2025, with $70 million in seed capital to bring ALM-based audio AI products to market. Was created by former researchers from Kyutai, DeepMind, and Meta who have been developing natural language supervision methods for training AI to understand audio data.
Private cloud for on-prem deployments available on enterprise plans. Also available through the AWS Marketplace as a pre-trained SageMaker model (gradium-tts-202512) for cloud-based scalability.
Produces the highest Elo score ratings in blind human evaluations for multiple languages. Provides high-fidelity voice output, including preservation of micro-trait characteristics (such as vocal fry, and breathiness), and produces natural-sounding results, including the ability to switch between languages mid-sentence.
Compared to ElevenLabs, Speechify, Hume AI, Speechmatics. Strengths include real-time bi-directional conversation support and semantic voice activity detection.
CEO demonstrates live voice cloning capabilities to thousands of people. Featured in Slator, and SiliconANGLE, for Gradium’s contributions to voice AI and the company’s recent funding.
| Region | Languages | Dialects | SSML Support | Number/Date Handling |
|---|---|---|---|---|
| Europe | English, French, German, Spanish, Portuguese | Regional variants | Yes | Full (localized formats) |
| Americas | English, Spanish, Portuguese | Regional variants | Yes | Full (USD, etc.) |
| Global Support | 5 | Multiple locales | Yes | Comprehensive |
For Realtime Applications - Streaming Inference
For Most Web Applications - Standard HTTP Endpoints
Provides an Asynchronous Client Library
High Performance Client
Compatible with Agent Frameworks
Compatible with Agent Frameworks
Deployed as a Managed Service
Can be Deployed On-Premise
Audio Generation from Text using Speech Synthesis Markup Language
Generates Real-Time Audio from Text
