On November 5, Gradium published a technical account of Phonon, its family of models that turn text into speech on the user's device. The weights had been released on October 20. The November article describes the runtimes, the differences between model variants, and evaluations of speed and intelligibility. Generation can take place offline, including in a compatible browser, after the application has the files it needs for the model and voice.
Which languages and licenses the Phonon family offers
Phonon 31m has 31 million parameters and synthesizes speech in English. Its model card lists a 64.5 MB eight-bit weights file and the Apache 2.0 license, which permits commercial use under its terms. Phonon 40m Multilingual uses one set of weights for English, French, German, Spanish, and Portuguese. Its file in the same format is 74.6 MB. Gradium also offers specialized 40m variants for each of the five languages.
The 40m variants are published under CC BY-NC 4.0 for noncommercial use. Gradium offers separate licensing for commercial products that use one of those variants. Access to the multilingual model files through the public repository requires accepting the conditions from an account. For an application intended for a business in Romania, the working language is a selection criterion. The multilingual model lists the five languages above, and Romanian is outside its documented coverage.
The published sizes refer to the quantized weights files. An application also includes the model runtime, the components that prepare the text, one or more voice files, and the memory required during generation. Gradium describes the output as mono audio at 24 kHz. For an application that distributes the model directly in a browser, the initial download size and the ability to retain the files on the device affect the experience separately from subsequent playback.
How playback starts before the whole sentence is generated
Phonon is a speech synthesis model. It takes text and produces sound, while speech recognition or translation can be added by other application components. The model produces intermediate audio representations in successive frames. A decoder turns them into playable sound. The runtime can deliver chunks of audio as it generates them, so playback can begin while the rest of the input is still being processed.
In the documentation for its own runtime, Gradium describes normalizing text before synthesis. Numbers, dates, currencies, and other symbols are prepared for pronunciation according to the selected language. The format matters for status messages, instructions, or reading application content aloud, where an order number and a calendar date have different pronunciation rules. The reference implementation in Transformers has a different preprocessing path, described separately in the model card.
Running on a processor, accelerator, or in a browser
Gradium built the ptts runtime to run on several kinds of hardware. Documented options include Intel and Arm CPUs, WebAssembly, GPU acceleration through CUDA, Vulkan, Metal, or WebGPU, and neural acceleration options through Apple Core ML and Snapdragon. The availability of each path depends on the system and package in use. The browser demo shows how audio can play inside a user's tab with generation taking place on the device.
The PyTorch reference implementation is available in the Transformers library from version 5.21, according to the technical account. Gradium describes it as an option for batch generation on GPUs, while its own runtime targets individual streams on the device. That distinction matters for a business comparing an internal tool that prepares many audio files with an interactive app that needs to start playback as soon as a user requests a sentence.
For the included voices, the application can use locally stored files. Gradium separately offers an online service for designing custom voices. Once a suitable voice has been obtained and included, local playback uses the model and voice available in the application. In a web app, offline mode also requires the page and its necessary resources to be available on the device already. How an application stores its own data remains an implementation decision.
The model card describes accessibility applications, local assistants, and reading instructions aloud. In those workflows, text can come from the displayed page or an internal application, and Phonon handles the playback stage. For an app that reads long documents, the documentation specifies that input is split into sentences. The model generates each segment in succession, with intonation determined at the segment level.
What the published evaluations measured
In the technical account, Gradium discusses Open TTS Leaderboard evaluations for intelligibility, speed, and naturalness. Intelligibility measures how faithfully the spoken words can be recognized. Time to first audio measures when playback can begin. Speed relative to the duration of the audio answers a different question about the pace of generation after startup. Gradium reports leading results for German and Spanish intelligibility and for time to first audio in the CPU streaming test.
These results are based on the models, languages, voices, and sets of sentences used in the evaluation. In an application, the time between pressing a button and hearing playback also includes starting the runtime, loading the voice, and preparing the text. A workstation that already has the model in memory and a browser opening the feature for the first time have different startup times. The documentation makes it possible to identify each stage of the playback workflow.
