July 2026 Summaries

A new Gradium TTS model is now available in public beta, enhancing pronunciation accuracy for complex cases such as phone numbers, email addresses, IBAN numbers, and time expressions, in several languages. This beta model, which requires no pre-processing or text normalization, aims to simplify life for developers by allowing text to be sent directly from databases or language models while ensuring accurate readings. The beta is accessible exclusively through the API, with users encouraged to provide feedback that will help refine the model based on real-world production cases. Participants who contribute feedback during the beta phase receive 1 million Gradium credits, and existing voices, including custom ones, remain compatible.

Speech-to-Text models excel at transcribing everyday speech but often struggle with rare, specific terms, such as newly introduced products or unique names, as they tend to default to more common, similar-sounding words. Keyword boosting addresses this issue by allowing users to provide a short dictionary of important terms, enhancing the model's accuracy by increasing the likelihood of correctly transcribing these words in real-time without retraining. This technique proves particularly useful in domains with specialized vocabulary, such as sports commentary or medical terms, where the model's baseline performance might falter. The boost parameter, adjustable in the model setup, determines the strength of preference for the listed keywords, with a typical recommended value of 3, which effectively recovers rare vocabulary while minimizing side effects. Users can implement keyword boosting via a simple JSON configuration in the model's WebSocket setup, making it a valuable tool for improving transcription accuracy in scenarios where specific terms are crucial.

Phonon is a 100-million-parameter on-device Text-To-Speech (TTS) model that excels in multilingual voice synthesis, supporting English, French, German, Spanish, and Portuguese. It demonstrates significantly lower word error rates and higher speaker similarity compared to other TTS models like NVIDIA's Magpie and NeuTTS Nano, with up to 3.5 times fewer errors in languages such as French, German, and Spanish. Designed to operate on a single CPU core with a memory footprint of around 200 MB, Phonon is optimized for efficient deployment on mobile devices, allowing for offline use in scenarios where cloud-based solutions are impractical. Its architecture, based on Continuous Audio Language Models with flow-matching for waveform generation, provides flexibility and resilience across different languages without the need for language-specific phonemizers, which contributes to its reduced error rates and enhanced performance. Phonon’s ability to clone a reference voice from a brief sample across all supported languages further distinguishes it from competitors, offering a scalable solution for privacy-sensitive and latency-critical applications.

Gradium and Keenable have partnered to enhance real-time voice AI and web search by integrating Keenable's fast web retrieval capability into Gradium's open-source Gradbot agent framework, which facilitates natural conversations. This collaboration addresses the latency issues commonly experienced by voice agents when retrieving up-to-date or rare information from the web, enabling Gradbot to access live data without significant delays. Traditional search APIs often result in pauses that disrupt conversational flow, but Keenable's ability to return results in under 200 milliseconds allows for seamless integration into the agent's reasoning process, ensuring that responses are quick enough to maintain a natural dialogue. This integration offers an improved experience where users can receive timely and accurate information, such as current headlines or train schedules, without noticeable pauses, transforming voice interactions from mere query responses to engaging conversations. The public can experience this innovation through a hosted demo, and integration documentation is available for those interested in incorporating this technology into their own systems.

Gradium, a company founded in September 2025 by the researchers behind Kyutai, has quickly established itself as a prominent player in real-time voice AI, having expanded its funding to $100 million just seven months post-launch with new investors like NVIDIA. This funding is set to bolster Gradium's AI research, product development, and international expansion, including a new office in the San Francisco Bay Area. Gradium's foundational infrastructure for real-time voice AI empowers developers and enterprises to create natural voice experiences through innovations in speech-to-text, text-to-speech, and conversational intelligence. Recent advancements include an improved Text-to-Speech model for complex content, semantic turn detection for more natural conversations, and tools like Gradium Translate and GradBot that facilitate seamless speech translation and voice agent development. These efforts underscore Gradium's mission to become a leader in voice-native AI applications, as it continues to attract enterprise customers across various industries and anticipates voice becoming the primary interface between people and intelligent systems.

RMC BFM Drive, powered by Gradium, offers a groundbreaking personalized radio experience in Renault's connected vehicles, utilizing AI to generate unique audio programs tailored to individual drivers. Launched in March 2026, this service allows users to input preferences such as topics of interest, trip duration, and preferred journalist voices, which the AI then uses to curate a dynamic audio sequence drawn from RMC BFM's extensive content library. This innovation is made possible by Gradium's advanced Text-To-Speech models that provide seamless, real-time voice synthesis, ensuring that the listening experience feels like a live broadcast with high-quality Pro Voice Clones preserving the distinct vocal characteristics of RMC BFM journalists. RMC BFM Drive is available for free through the openR link Play Store in Renault vehicles, exemplifying the collaboration between RMC BFM Group, a major French media company, and Renault Group, France's leading automotive brand, with Gradium providing the sophisticated AI infrastructure behind the service.