Hugging Face Phonon TTS: 65MB On-Device Voice Tops Benchmarks
By Ali Al-Montaser — Senior AI Technical Editor
Hugging Face launched Phonon TTS, a 65MB and 75MB on-device text-to-speech model supporting five languages without servers or network. It ranked first in German and Spanish intelligibility and led CPU time-to-first-audio, surpassing 200,000 downloads in two weeks under an Apache 2.0 license for the 31m variant.
Executive Overview
Hugging Face has officially announced Phonon TTS, a text-to-speech model that runs entirely on-device without a server or network connection. Available in 65 MB and 75 MB variants and supporting five languages, Phonon surpassed 200,000 downloads in its first two weeks. The Phonon 31m model is released under the Apache 2.0 license for free commercial use, while other models fall under the non-commercial CC BY-NC license. The model ranked first in German and Spanish intelligibility and led the CPU streaming time-to-first-audio benchmark, competing against models approximately 100 times larger.
📊 Official Technical Specifications & Data Sheet
| Technical Axis | Confirmed Official Data |
|---|---|
| 💰 Pricing & Usage Cost | Phonon 31m is completely free under Apache 2.0 (including commercial use). Phonon 40m is under CC BY-NC for non-commercial use. Commercial use of CC BY-NC models is priced by number of users, not by volume of generated speech. |
| 🌐 Platforms & Immediate Availability | Hugging Face Hub (direct download), Transformers from v5.21, dedicated ptts framework. Runs on phones, laptops, single-board computers, and inside a browser tab. More than 10 backends: CPU (Intel, Arm, WebAssembly), GPU (CUDA, Vulkan, Metal, WebGPU), NPU (Apple Core ML, Snapdragon). |
| ⚡ Performance & Speed Benchmarks | First place in German and Spanish intelligibility, and within the top three in every evaluated language. Leads the CPU streaming time-to-first-audio benchmark. Competes with models roughly 100 times larger. Over 200,000 downloads in the first two weeks. |
| 🛡️ Security & Breach Resistance | Fully local on-device processing: text never leaves the user's device and is not sent to any server. No network connection required, eliminating the network attack surface and achieving full data privacy. |
| 🧠 Memory Context (Context Window) | Model size 65 MB and 75 MB (very small memory footprint suitable for edge devices). Uses autoregressive generation of audio latents with flow matching. |
| 🌍 Arabic & Regional Support | Arabic is not supported in the current release. The five languages are English, French, German, Spanish, and Portuguese. Arabic training is available on request through design partners. |
Deep-Dive Features & Architecture
Phonon belongs to the Pocket TTS lineage and uses flow matching for audio generation, a fundamentally different approach from the modeling techniques in other TTS systems, enabling high quality at an extremely small size. The team faced challenges in early numerical training stability and required several technical tricks to produce high-quality results across all five languages. The model is available in 65 MB and 75 MB sizes and includes more than 100 built-in voices, with the ability to create custom voices via Gradium Studio through cloning from a short recording or text-based description. Phonon handles cases where TTS systems typically struggle, such as phone numbers, dates, and symbols.
The company developed a new inference framework called ptts dedicated to Phonon, supporting more than 10 backends across CPU, GPU, and NPU to achieve the best performance on the widest range of devices. Hugging Face also added Transformers v5.21 integration as a reference PyTorch implementation, targeting GPU productivity through batched generation and full-graph torch.compile, making it a foundation for building large-scale audio data generation systems. The model comes as a modular layer on top of Pocket TTS with approximately 200 lines describing only the differences.
The launch included applied demonstrations from the team and partners: an Apple Watch without Wi-Fi explaining coffee preparation step by step, an iPhone without signal counting exercise reps, an Autonomous Lamp robot speaking with personality, an offline dinosaur game turned into a radio program, a sci-fi game with local voice, and reading Proust page by page. Developers have also begun building on it: an ExecuTorch conversion runs Phonon on the guest GPU in the browser, Micdrop brings it to Node.js applications, and Spaces like smallest-s2s.
Benchmark & Competitive Performance
On the Open TTS Leaderboard, a new Hugging Face project featuring more than 9,000 TTS models on the platform, models are evaluated across 23 languages on three axes: intelligibility (Word Error Rate, WER), speed, and quality. Phonon achieved first place in intelligibility for German and Spanish and entered the top three models in every evaluated language, competing against models roughly 100 times larger. It also led the standard CPU streaming time-to-first-audio benchmark, a critical metric for real-time applications on edge devices.
Industry Impact & Enterprise Adoption
Phonon's on-device architecture eliminates the need for network connectivity and server infrastructure, enabling voice AI in privacy-sensitive environments such as healthcare, finance, and personal devices. The Apache 2.0 license for Phonon 31m lowers the barrier for commercial adoption, while the CC BY-NC license for larger models offers a path for non-commercial research. With over 200,000 downloads in two weeks and a growing ecosystem of integrations (ExecuTorch, Micdrop, Spaces), Phonon is positioned to accelerate the deployment of local voice interfaces across mobile, embedded, and browser platforms. The availability of Arabic training through design partners signals future expansion into new markets.
Conclusion
Hugging Face's Phonon TTS represents a significant leap in efficient, private, and accessible speech synthesis. By delivering competitive quality at 65–75 MB and leading key benchmarks, it challenges the assumption that high-quality TTS requires massive models and cloud infrastructure. As the ecosystem grows and language support expands, Phonon could become a foundational technology for on-device voice AI.
Media Source: Hugging Face | البيان الرسمي للشركة: المصدر الأصلي | Fact Verification & Analysis: AI Tools Oasis
Frequently Asked Questions
Phonon 31m is completely free under the Apache 2.0 license, including commercial use. Other models (Phonon 40m) are licensed under CC BY-NC for non-commercial use; commercial use is priced based on the number of users, not the volume of generated speech.
Phonon does not currently support Arabic in the announced release. The five officially supported languages are English, French, German, Spanish, and Portuguese. The company announced that training for Arabic, Korean, and Japanese is available on request through design partners.
Phonon is available on the Hugging Face Hub and runs locally on phones, laptops, single-board computers, and inside a browser tab. The ptts framework supports more than 10 backends: CPU (Intel, Arm, WebAssembly), GPU (CUDA, Vulkan, Metal, WebGPU), and NPU (Apple Core ML, Snapdragon). It is also available via Transformers from v5.21.
Phonon is available in two sizes: 65 MB and 75 MB. It includes more than 100 built-in voices, with the ability to create custom voices via Gradium Studio using voice cloning or text-described voice design.
Phonon ranked first in intelligibility for German and Spanish and placed among the top three models in every evaluated language, competing against models roughly 100 times larger. It also led the standard CPU streaming time-to-first-audio benchmark.
AI Tools Oasis Team
Bringing you the latest news and analysis in the world of Artificial Intelligence with accuracy and credibility. Follow us for all updates.
