Recently, Ollama, the open-source platform for running large models locally, announced that its Mac app has officially received a major update, adding text-to-speech for the first time. This upgrade marks a full leap for local AI assistants from the traditional "can read and write" stage to an interaction stage where they "can speak," with Windows and Linux versions set to follow.
On the underlying technology and privacy protection front, the first supported speech model is the Phonon series developed by Gradium. The model is compact and technically advanced. Users do not need to rely on any cloud services and can download it directly through Ollama and run it on Apple Silicon Macs, making it a great fit for real-world scenarios that prioritize data privacy and offline use. Leveraging macOS, Ollama uses Apple's open-source machine learning framework MLX, fully tapping into the hardware advantages of Apple Silicon's unified memory architecture to significantly improve runtime efficiency.
Official test data shows that when running Phonon on a MacBook Pro with an M5 Max chip, its speech generation speed is 35% faster than Gradium's own implementation, while maintaining an extremely low word error rate comparable to the Hugging Face Transformers reference implementation. In addition, the feature supports real-time speech generation. Users do not have to wait for the AI to output the entire text before they can hear the corresponding speech in sync as the response is generated, delivering a more natural and fluent human-like conversation experience.
In terms of specific voices and language support, Phonon currently includes separate models for five languages: English, French, German, Spanish, and Portuguese, and also offers a multilingual general-purpose model that supports all five languages. It also allows users to freely choose different accents, such as American or British, according to their preference. As a popular open-source project on GitHub, Ollama currently has nearly 185,000 stars and about 18,600 forks, continuing to provide developers and everyday users with a convenient and secure solution for running large models locally.