Hugging Face’s Open TTS Leaderboard has been updated with the release of Phonon, a new suite of on-device text-to-speech (TTS) models developed by Gradium. Phonon has quickly established itself at the forefront of TTS technology, particularly for its high-quality output and efficient performance on resource-constrained hardware. The models have garnered significant attention, achieving over 200,000 downloads within their first two weeks of release.
According to details received by The Chenab Times, Phonon models have secured top rankings across various evaluations. They notably place first in German and Spanish intelligibility and lead the CPU streaming benchmark for time-to-first audio, indicating rapid speech generation startup. Phonon also ranks within the top three in every language evaluated, competing effectively against models that are up to 100 times larger. This performance is attributed to its architecture, which is built for local inference, ensuring that text data remains on the device and enabling offline voice functionality.
Phonon is available in two main sizes: Phonon 31m and Phonon 40m. The Phonon 31m model, which speaks English, is released under the Apache 2.0 license, making it free for commercial use. The Phonon 40m model, a multilingual variant covering English, French, German, Spanish, and Portuguese, is available under a non-commercial license (CC BY-NC 4.0), with commercial use subject to pricing based on the number of users. Both models ship with over 100 built-in voices and are designed to run without requiring a server or network connection.
The development team at Gradium has emphasized Phonon’s ability to deliver natural and accurate speech on devices such as phones, laptops, and single-board computers. Phonon’s smaller parameter count, around 31 million or 40 million, allows for exceptionally fast inference times. For instance, Phonon 31m can start generating speech in as little as 7 milliseconds on a MacBook Air M5 CPU, a significant improvement over other models like Kokoro, which can take over 500 milliseconds. On a Raspberry Pi 5, Phonon 31m starts speaking in 38 milliseconds and operates at 3.7 times faster than real-time.
Phonon’s capabilities extend to voice cloning, allowing it to reproduce a reference speaker’s tone, accent, and cadence from a sample as short as 10 seconds. This feature is crucial for applications requiring a consistent or personalized voice experience, such as navigation apps, language learning tools, and accessibility software designed to preserve a user’s original voice.
The models have been integrated into various popular platforms and libraries, including Hugging Face Transformers, MLX-Audio for Apple Silicon, LM Studio, and Ollama, demonstrating their broad applicability. Gradium also provides inference code and backends for multiple platforms, including CPU, Metal, CUDA, Vulkan, WebGPU, and Apple Core ML, facilitating integration into diverse applications.
Benchmarking results further highlight Phonon’s competitive edge. On the Seed-TTS English benchmark, Phonon has achieved a word error rate (WER) as low as 1.00%, outperforming larger models. In multilingual evaluations, Phonon averaged a 1.07% WER across French, German, and Spanish, significantly better than NVIDIA’s Magpie TTS, which recorded 2.26% WER. These metrics, alongside human preference listening tests where Phonon ranked first, underscore its advanced speech quality and naturalness.
The Phonon models are designed to address the limitations of traditional cloud-based TTS solutions, which can incur per-request costs, require constant network connectivity, and raise privacy concerns when sensitive text data must be transmitted. By operating entirely on the user’s device, Phonon offers unlimited usage, offline functionality, and enhanced data privacy, making it a compelling option for a wide range of consumer and enterprise applications.
❤️ Support Independent Journalism
Your contribution keeps our reporting free, fearless, and accessible to everyone.
Or make a one-time donation
Secure via Razorpay • 12 monthly payments • Cancel anytime before next cycle
(We don't allow anyone to copy content. For Copyright or Use of Content related questions, visit here.)
The Chenab Times News Desk
