April 2026 Summaries

The text delves into the evaluation of text-to-speech (TTS) models, emphasizing the importance of pronunciation accuracy, which is typically measured by comparing transcribed audio with the original text input. It highlights the challenges of normalization in TTS, such as dealing with homophones and complex numbers, and introduces popular tools like the Whisper English normalizer and Kyutai's tts_longeval French normalizer to aid this process. The text also discusses the Word Error Rate (WER) as a measure of TTS performance, using the MiniMax Multilingual TTS Test Set for benchmarking, and compares the performance of various TTS models, with Gradium emerging as a leader in several languages. Despite the focus on WER, the text argues that a comprehensive TTS system must also ensure low latency, natural-sounding voices, and robust handling of complex linguistic features. It concludes by suggesting that while current benchmarks have reached a saturation point for simpler tasks, future progress in TTS will require more challenging tests and broader measures of robustness.

Gradium, a voice AI platform, is available as a managed SaaS subscription on AWS Marketplace and as a deployable Amazon SageMaker image for in-VPC deployment, catering to different organizational needs. The AWS Marketplace option offers a fully managed service with billing through an AWS account and no need for infrastructure management, providing 45 million credits per month for TTS and STT capabilities in five languages. This path is ideal for teams without stringent data residency requirements. Alternatively, the SageMaker image allows organizations to deploy the TTS model within their private AWS environment, maintaining data boundaries necessary for compliance with strict policies like HIPAA. This option is suitable for teams that prefer managing their own infrastructure and need to ensure that audio data does not leave their VPC. The choice between these options hinges primarily on data processing constraints, with the Marketplace plan requiring less overhead and the SageMaker image offering greater control over data handling.

Gradium launched Gradbot, an open-source framework designed for rapid prototyping of voice agents, in December 2025, building on its high-quality Text-to-Speech and Speech-to-Text models. Gradbot enables developers to create voice experiences with minimal code, using a Rust-based multiplexing engine that synchronizes speech-to-text, language model inference, and text-to-speech in real time, while managing conversational flow and handling interruptions naturally. It is particularly useful for applications like 3D NPC games, travel booking assistants, and various experimental interfaces, thanks to features like voice activity detection, silence management, and asynchronous tool calling that ensure smooth and engaging user interactions. Gradbot is ideal for prototyping innovative voice applications, allowing developers to focus on ideas without the burden of extensive infrastructure setup, while Gradium's models can be utilized for more reliable, scalable production environments through other orchestrators.

Phonon is Gradium's on-device text-to-speech (TTS) model, notable for its compact size of approximately 100 million parameters and its ability to run efficiently on a single CPU core, making it suitable for mobile devices and browser applications without requiring GPU acceleration. In evaluations using the Seed-TTS English benchmark, Phonon achieved a word error rate (WER) of 1.48% and a speaker similarity of 56.37%, outperforming larger models like Kani-TTS2 and NeuTTS Air. The model supports voice cloning from a 10-second sample, allowing it to replicate the tone, accent, and cadence of the reference speaker without needing a transcription of the reference audio. By leveraging efficient audio language models, Phonon offers high-fidelity text-to-speech capabilities, making it applicable for various uses, such as accessibility apps preserving user voices, language learning tools maintaining consistent teacher voices, and consumer apps offering diverse voice styles. The evaluation process for Phonon involves transcribing generated speech back to text to compute WER and using speaker embeddings to measure speaker similarity, with the model currently in private beta and under active development for further improvements.

Gradium offers two distinct text-to-speech solutions: the cloud-based Gradium API and the on-device Gradium Phonon model, each tailored for different use cases. The Gradium API provides a comprehensive, flexible service capable of handling any language and voice, ideal for real-time applications and environments where network access is available. It operates on a variable cost model, which can become unsustainable at high volumes. In contrast, Gradium Phonon is an on-device model designed for specific, well-defined use cases that require offline functionality, unlimited scalability, and adherence to data privacy standards. It is optimized for a single voice and language, offering a fixed-cost model per device that eliminates network dependency, making it suitable for low-connectivity environments and privacy-sensitive applications. Both solutions can complement each other within the same application, allowing developers to choose the best tool based on connectivity, cost, and quality needs. Currently, Gradium Phonon is in private beta, inviting select partners to develop tailored models for their specific voice and device requirements.