TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Hugging Face says Phonon is a compact text-to-speech model that generates speech locally, without a server or network connection. The company reports strong results on selected Open TTS Leaderboard measures, including first place for intelligibility in German and Spanish, but the model sizes have different licenses and the results are not an independent evaluation.
Hugging Face says it has released Phonon, a compact text-to-speech model that can generate speech on a user’s device without a server or internet connection. The release matters for developers building offline voice features and for users seeking to keep text on their own devices; Hugging Face also reports that Phonon ranked first for intelligibility in German and Spanish on its Open TTS Leaderboard.
Phonon is available in two model sizes, approximately 65 MB and 75 MB, and supports English, French, German, Spanish and Portuguese across the release. Hugging Face says it includes more than 100 built-in voices and can handle formats such as phone numbers, dates and codes. Users can also create custom voices through voice cloning and voice design in Gradium Studio.
The two models have different licensing terms. Phonon 31m, which speaks English, is released under Apache 2.0 and can be used in commercial products. Phonon 40m, which covers all five languages, is under a non-commercial license. Hugging Face says commercial use of the other models is priced according to the number of users rather than the amount of speech generated.
For evaluation, Hugging Face says Phonon ranked first in German and Spanish for intelligibility and placed in the top three in every language it evaluated. The company also says the models led its CPU streaming benchmark for time-to-first audio. These are claims about particular leaderboard measures, not a claim that Phonon leads every category or every TTS system. The leaderboard assesses factors including intelligibility, speed and naturalness, and the source report does not provide a full independent assessment of the results.
Offline Speech for Smaller Devices
Running speech generation locally can reduce dependence on cloud services. According to Hugging Face, Phonon needs no network connection, so text sent to the model can remain on the device. This could be useful where connectivity is poor, where an application needs to respond without a server round trip, or where developers want to limit the amount of user text sent elsewhere. The privacy benefit depends on the implementation: the company describes local processing, but the report does not audit every app or integration built on the model.
The model’s small size and reported CPU speed could also make speech features more practical on devices with limited resources, including phones, laptops and single-board computers. Hugging Face points to demonstrations running on an Apple Watch with Wi-Fi off and an iPhone with no signal, as well as browser and game projects. These examples show intended uses, but they are not a substitute for published, device-by-device measurements of battery use, latency or audio quality.
Licensing affects who can adopt the release. The English-only Apache 2.0 model offers a permissive route for commercial development, while the five-language model’s non-commercial terms constrain commercial use without a separate arrangement. Developers need to check the specific model’s license before deploying it.
Products Worth Considering
Bose QuietComfort Ultra Earbuds (2nd Gen) – Wireless Noise‑Cancelling Earbuds with Mic, Immersive Audio, USB‑C Charging, Up to 6 Hours Batte
Apple iPad mini (A17 Pro), 128GB, Wi-Fi 6E, Space Gray
Apple MacBook Air 15-inch with M5 chip, 16GB memory, 512GB SSD, and Liquid Retina display (Midnight)
As an affiliate, we earn on qualifying purchases.
From Pocket TTS to Phonon
Hugging Face says Phonon builds on the Pocket TTS lineage and uses flow matching to generate audio, a different approach from those used by many other text-to-speech systems. The company says this approach helped it pursue audio quality in a small model, while also creating training challenges, including numerical instability. It reports using techniques to stabilize training and produce results across five languages.
The company says it announced Phonon in April and spent the next six months developing it with partners and early users. That work led to smaller models, additional languages and support for custom voices. Hugging Face reports that Phonon received more than 200,000 downloads in its first two weeks. That figure is a download count for the stated period; it does not establish how many people used the model or how often.
The release includes an inference framework called ptts, which Hugging Face says supports more than 10 backends across CPU, GPU and neural processing unit hardware. A reference implementation is also available through Transformers, from version 5.21, according to the report. The company describes ptts as aimed at single-stream local inference and Transformers as supporting GPU batch generation and related development work.
“Your text stays there and the voice works without a network.”
— Hugging Face, in its Phonon launch report
Limits of the Reported Results
The report does not give enough detail to independently verify the leaderboard comparisons, including the complete test setup, model-by-model results and the specific language set used for every claim. Hugging Face describes the Open TTS Leaderboard as a new project, and its published measures cover distinct aspects of performance; a strong ranking in one measure does not establish overall superiority.
Practical performance will also depend on the device, runtime and application. The report does not provide comparable figures for battery consumption, memory use, sustained generation speed or quality across different hardware. It also does not establish how custom voice creation works in every deployment or what protections apply to voice samples. The download count does not show how many downloads became active use.
License terms require attention: only the 31m English model is identified as Apache 2.0, while the multilingual 40m model is non-commercial. The report does not set out all terms for commercial access to the other models, beyond saying pricing is based on user count. Developers should consult the relevant model and service terms.
Developer Integrations and Testing
Developers can test Phonon through Hugging Face’s demos and integrations, including Transformers and the ptts inference framework. The company also cites projects using an ExecuTorch conversion in a browser, a Node.js integration called Micdrop, and community Spaces. The next practical test is whether independent developers can reproduce the reported speed and quality across the range of supported devices.
Readers evaluating the model should compare the exact model, language, license and runtime they plan to use. Further leaderboard results and third-party measurements could clarify how Phonon performs beyond the reported intelligibility and CPU streaming results. Hugging Face’s report does not announce a date for additional evaluations or a broader commercial release.
Key Questions
What is Phonon TTS?
Phonon is a text-to-speech model from Hugging Face designed to generate speech locally on devices such as phones and computers, without requiring a network connection for inference.
Which languages does it support?
The release covers English, French, German, Spanish and Portuguese. Hugging Face identifies the 31m model as English-only and the 40m model as multilingual.
Can businesses use Phonon commercially?
Hugging Face says the English-speaking Phonon 31m model is Apache 2.0 and available for commercial use. The 40m model is under a non-commercial license; developers should review the applicable terms before commercial deployment.
What does the leaderboard ranking show?
Hugging Face reports that Phonon ranked first for intelligibility in German and Spanish, placed in the top three in all languages it evaluated, and led its CPU streaming benchmark for time-to-first audio. These findings concern specific measures and do not establish that it is best in every TTS category.
How many downloads did Phonon receive?
Hugging Face says the models were downloaded more than 200,000 times in their first two weeks. The report does not say how many individual users or active installations that represents.
Source: huggingface.co
