The Pulse
Gradium Says Its On-Device Phonon TTS Ranks First in German and Spanish Intelligibility on Hugging Face's Open TTS Leaderboard
Gradium says its Phonon TTS runs on-device without a server or network and comes in two models covering five languages. Gradium says Hugging Face’s Open TTS Leaderboard ranks Phonon first in German and Spanish intelligibility, and Phonon 31
AI.info Team · Written with AI · fact-checked ·
Gradium’s Phonon TTS targets a trade-off its creators describe directly: smaller speech models can sound worse, while models that sound good may not fit on a small device. Gradium says Phonon runs on a user’s device without a server or network, and that it addressed both constraints.
Two models, five languages
Phonon comes in two sizes, about 65 MB and 75 MB, and together the models support English, French, German, Spanish and Portuguese. Gradium says they offer more than 100 built-in voices, and users can create voices through Gradium Studio. Gradium says Phonon handles details such as phone numbers, dates and codes, which the post says most other text-to-speech systems often stumble on.
Gradium says Phonon runs on phones, laptops, single-board computers and in a browser tab. The English-only Phonon 31m model uses the Apache 2.0 license, allowing commercial use; Phonon 40m covers all five languages under a non-commercial license. The post says commercial pricing for the other models is based on the number of users, not the amount of speech generated.
Gradium says it announced Phonon in April and spent the following six months building on it with partners and early users. Their feedback helped shape the release: the models got smaller, and languages and support for custom voices through voice cloning and voice design were added. The model follows the Pocket TTS lineage and uses flow matching to generate audio, which the post says allowed Gradium to bring this level of quality to a small model.
Leaderboard results, with Phonon 31m en at 28 ms (95th-percentile) to first audio in the CPU streaming benchmark
Hugging Face’s Open TTS Leaderboard compares open models across 23 languages, measuring intelligibility, speed and naturalness. Its intelligibility measures include word error rate; speed measures include inverse real-time factor and time to first audio. For naturalness, the leaderboard uses UTMOS for English and KOMOS for Korean, and it evaluates both fixed voices and voice cloning.
Gradium says Phonon ranks first for intelligibility in German and Spanish and in the top three in every language the team evaluated, next to models up to 100 times its size. In the CPU streaming benchmark, Phonon 31m English has a 95th-percentile time to first audio of 28 ms, while Phonon 40m English records 38 ms; Pocket TTS records 165 ms.
More than 10 inference backends
Gradium built ptts, an inference framework for Phonon, to get the best performance on a wide variety of devices. Gradium says it supports more than 10 backends, including Intel, Arm and WebAssembly on CPU; CUDA, Vulkan, Metal and WebGPU on GPU; and Apple Core ML and Snapdragon on NPU. Phonon is also available in Transformers from version 5.21 as its reference PyTorch implementation.
ptts runs a single stream on a device, while Transformers targets GPU throughput through batched generation and full-graph torch.compile. Being native PyTorch, the Transformers implementation is a base for integrating or studying the model. Its code builds on Pocket TTS, with about 200 lines describing only what differs in Phonon.
Offline demos, from a watch to a browser
The launch demos include an Apple Watch that talks through a coffee brew with Wi-Fi off and an iPhone that counts reps with no signal. Another demo turns an offline dino runner in a browser into a radio show hosted by Rex. The post also describes a sci-fi game voiced on the player’s own machine and Proust read aloud page by page.
The Hugging Face team built the Transformers integration, while developers have made an ExecuTorch conversion for running Phonon on a visitor’s GPU in a browser and Micdrop support for Node.js apps. Phonon was downloaded more than 200,000 times in its first two weeks, according to the post. Gradium says it also trains custom Phonon models with design partners, and that it can train for other languages or fine-tune on a product’s own vocabulary.