Gradium Open-Sources Phonon: A 31M-Parameter On-Device TTS Model

What Phonon Is: Squeezing TTS into 64.5 MB

According to MarkTechPost, Gradium has open-sourced the Phonon family of on-device text-to-speech (TTS) models. The English version, Phonon 31m, is released under the Apache 2.0 license, has a model size of just 64.5 MB, and takes about 7 milliseconds to start producing sound on a laptop CPU.

These numbers are worth unpacking. 31M parameters means it's roughly one-hundredth the size of mainstream cloud TTS models; at 64.5 MB, it can fit inside mobile apps or even embedded devices; and a 7 ms first-audio latency approaches a "press-and-speak" interaction experience. Together, these three point to one conclusion: Phonon's goal isn't extreme audio quality, but turning speech synthesis into a local capability that can be distributed along with the device.

Engineering Trade-offs of On-Device TTS

Why Latency and Size Matter More Than Audio Quality

The typical cloud TTS pipeline is: upload the recording, run inference on the server, send the audio back. Every step introduces network jitter, and the whole thing simply fails under weak or offline conditions. On-device approaches eliminate all of these steps, at the cost of limited model capacity. Phonon's choice to work within 31M parameters is essentially trading acceptable audio quality for deterministic response speed and privacy guarantees—voice data never leaves the device.

A TTS model's "first-audio latency" (Time to First Byte / First Audio) is not the same as end-to-end inference latency. The former refers to the time from input text to the first audio sample coming out of the speaker, including text front-end processing (tokenization, phonemization, prosody prediction) plus the first-frame inference of the acoustic model and vocoder; the latter also counts the time to generate the complete audio. Cloud pipeline latency comes mainly from network round-trip time (RTT) and queuing, whereas on-device latency is almost entirely determined by model compute. So the 7 ms figure is typically a best-case measurement taken on a warmed-up model, with short text, on a specific CPU—real-world experience will vary with text length and hardware. Understanding this distinction helps in judging whether a vendor's advertised latency numbers are reproducible in a real product.

What Apache 2.0 Actually Means

Apache 2.0 is a business-friendly license that permits modification, distribution, and commercial use, requiring only that copyright notices and the license text be retained. For product teams hoping to integrate offline speech into an app, this has far more practical value than research-oriented open-source licenses.

Among open-source AI models, license differences have a big impact on adoption. Common research licenses (such as the CC-BY-NC family or custom "research-only" terms) prohibit commercial use or impose restrictions on user scale or domain; Apache 2.0, by contrast, is a standard permissive license recognized by the OSI that, apart from requiring retention of copyright and license notices and including a patent grant clause, places no restrictions on commercial use, modification, or redistribution. Two caveats: first, the license covers only the model weights and code themselves—the licensing of training data and voice/timbre assets must be confirmed separately; second, if a product involves voice cloning or a specific speaker's timbre, it may also run into compliance requirements around personality rights and labeling of synthetic speech. So "commercially usable" is a necessary but not sufficient condition.

What It Means for Developers

Phonon fills in a relatively weak link in the on-device AI stack. Vision and language models already offer plenty of small-parameter options, but high-quality, commercially usable local TTS options have long been limited. With 7 ms-class latency and a 64 MB footprint, it suits scenarios such as accessibility reading, offline navigation announcements, and voice feedback for smart hardware.

What needs to be viewed soberly is that the original information did not disclose audio quality benchmark data, the number of supported languages, inference framework dependencies, or hardware compatibility range. The upper limits of 31M parameters for multilingual and emotionally expressive speech still need to be verified in practice. For most products, Phonon is better suited as an "offline fallback" or "low-latency first choice" rather than a full replacement for high-fidelity cloud synthesis.

On-device TTS typically uses a two-stage architecture: an acoustic model (such as a lightweight Transformer- or convolution-based network) maps text features to an intermediate representation like a mel spectrogram, and a vocoder then reconstructs the spectrogram into a waveform. The bulk of a model's size often lies in the vocoder and embedding tables; a scale of 31M parameters and 64 MB means both have been significantly compressed, possibly using parameter quantization (e.g., int8), knowledge distillation, or a lightweight vocoder design. Such designs place requirements on the inference framework: common deployment paths include ONNX Runtime, NCNN, or vendors' in-house mobile inference engines, and differences in operator support and quantization precision across frameworks directly affect actual latency and audio quality. So when choosing, beyond looking at parameter count, you also need to confirm the officially provided runtime bindings, target platforms (ARM/x86), and whether streaming chunked synthesis is supported.

In a Nutshell

Phonon 31m turns TTS into an embeddable, commercially usable, millisecond-responsive local component, lowering the barrier to on-device voice interaction another notch.