PhononTranslation site

If you want to add a text-to-speech feature to your app, it used to be easy to start with a cloud-based speech API: it’s easy to integrate, but each synthesis requires an internet connection and involves sending text, which comes with costs and depends on service availability. Gradium’s Phonon offers an alternative approach:Download the speech model to the device and convert text to speech locally.

It is particularly well-suited for developers focused on offline use, data boundaries, and edge deployment. However, lightweight does not mean without limitations; when choosing a version, you must consider the language, license, and entry point together.

What is a phonon?

Phonon is a series of open-source, weight-based text-to-speech models released by Gradium on October 21, 2026, covering various languages and deployment methods. Its core selling point is not chatbot functionality, but rather the ability to convert existing text into playable speech locally, serving as the voice output layer for readers, assistants, and application interfaces.

Gradium specializes in speech models and related infrastructure, and its team has a background in speech research at Kyutai. Its published materials mention CTO Laurent Mazaré; the company’s capabilities should be assessed based on its current models and documentation, and its value does not need to be demonstrated by unverified funding or team size.

First, let's take a look at the differences between the three versions

Chinese is not included among the languages listed here.Just because there are “multilingual” versions doesn’t mean it should be recommended as a Chinese dubbing tool. Nor can the conclusion regarding Apache 2.0 be applied to the entire product line based solely on the 31M English version.

What problems can it solve?

Phonon provides local synthesis capabilities for text-to-speech, English learning aids, desktop assistants, and device voice feedback. The model is compact, making it ideal for on-device deployment and experimentation; once the necessary resources are downloaded, local synthesis also reduces reliance on cloud requests.

The official documentation provides various integration paths, including options for Rust, Python, Swift, and browser-based solutions, covering backends such as CPU, NVIDIA, Apple, Vulkan, and WebGPU. However, the fact that “the project supports multiple devices” does not mean that every entry point supports all platforms: actual support depends on the selected backend, version, and example requirements.

Running locally can alter the flow of data, but it does not automatically ensure that the entire application operates offline. If the chat model described earlier still calls cloud APIs, or if the application retains telemetry and synchronization capabilities, the text may still pass through other services. When designing privacy-sensitive workflows, you should examine the entire data path.

What should you keep in mind when getting started with Ollama?

On October 24, 2026, Ollama announced support for voice models. The entry described in the announcement requires Ollama 0.40.11 or later, running on an Apple Silicon Mac; Windows and Linux are still listed in the announcement as part of the future support plan and should not be described as already available.

If you meet the requirements, follow the steps in the official example:

ollama pull phonon
ollama run phonon "Hello from Ollama."
ollama show phonon

The first two steps involve downloading the model and generating a sentence in English; the third step allows you to view the model information and available voice options. The official platform also provides options to specify a voice, export audio files, and integrate via API. For your first test, we recommend starting with short sentences, then testing numbers, proper nouns, punctuation, and longer paragraphs; do not judge the overall audio quality of an entire article based solely on a single demo sentence.

If you are using Python, Swift, or another standalone entry point, you should refer to the Phonon project documentation rather than simply adopting Ollama’s platform restrictions. When using voice mimicry or distributing audio, you must also verify the speaker’s authorization and the rights to the relevant material.

How does it compare to other local TTS services?

You can’t simply determine which one is better based on the number of parameters alone. We recommend comparing the first paragraph of the same text in terms of wait time, continuous reading, pronunciation of proper nouns, memory usage, and integration complexity. Sound quality and speed still need to be compared on the target device using actual text; official demos should not be directly used as a performance ranking.

Costs and Future Directions

Phonon's open-source model can be deployed independently under the applicable license, but there are still costs associated with hardware, storage, packaging and maintenance, and application development. Gradium's cloud service plans are not considered equivalent to the cost of an on-premises model, nor is there any guarantee that all audio files or future versions will be licensed under the same terms.

From the model’s release on October 21 to its integration into the Ollama voice workflow on October 24, one notable change is that the deployment process is becoming more straightforward. For developers of English-language applications, this is a viable option for building small-scale prototypes; however, for users primarily focused on Chinese voiceovers, it should not currently be at the top of their list.

Official Access Points and References

Phonon Product and Model Versions https://gradium.ai/phonon

Gradium Issues an Announcement https://gradium.ai/blog/gradium-releases-phonon

Team Introduction https://gradium.ai/about/who-we-are

Phonon Open Source Project https://github.com/gradium-ai/xn-ptts

Ollama Speech Model Announcement and Usage Guide https://registry.ollama.com/blog/speech-models

Basis for Comparison:Kokoro-82M Model Card,Pocket TTS Official Project https://huggingface.co/hexgrad/Kokoro-82M https://github.com/kyutai-labs/pocket-tts