Contents
If your app needs speech synthesis that works offline on users’ devices, Phonon is worth trying. Its small models deliver strong intelligibility and CPU streaming results, but licensing differs between models. The English model allows commercial use under Apache 2.0, while the multilingual model requires a separate commercial agreement.
Key Points#
- Phonon offers English and five-language models, with model packages of approximately 65 MB and 75 MB.
- According to Gradium, it ranked first for German and Spanish intelligibility and had the shortest time to first audio in the leaderboard’s CPU streaming benchmark.
- Phonon 31m uses Apache 2.0, while Phonon 40m uses a noncommercial license.
- Developers can use ptts for on-device inference or Transformers for PyTorch integration and batch generation on GPUs.
Choose the Right Model#
Phonon has two main options. Phonon 31m supports English and allows commercial use under Apache 2.0. Phonon 40m supports English, French, German, Spanish, and Portuguese and is available under a noncommercial license. According to Gradium, commercial pricing for models with usage restrictions is based on user count rather than the amount of speech generated.
The release includes more than 100 built-in voices. Gradium Studio also supports voice cloning from short recordings with the speaker’s consent, as well as voice design from text descriptions. The resulting voice files can be downloaded for local inference.
Japanese, Korean, and Arabic are not among the five languages supported by the standard multilingual version. Gradium offers custom training for additional languages and specialized terminology, but you will need to confirm availability, quality, and cost directly.
Read the Benchmarks Carefully#
In its November 5, 2026 launch report, Gradium states that Phonon ranked first for German and Spanish intelligibility and placed in the top three for every language evaluated. It also reports the best time to first audio in the Open TTS Leaderboard’s CPU streaming benchmark.
These results apply to specific conditions and do not prove that Phonon produces the best speech in every situation. Intelligibility measures whether spoken words can be recognized correctly. Time to first audio measures how quickly playback can begin. Neither metric alone guarantees natural delivery, accurate pronunciation of your specialized terminology, or consistent performance on long passages.
The leaderboard has a Listen tab for comparing outputs. Use it alongside the scores, and check the latest listings before deciding to adopt Phonon in your product. Rankings may change, and the cited report does not include precise latency figures or enough hardware information to predict on-device performance.
Choose an Inference Method#
Use ptts if your main requirement is running a single audio stream locally. Gradium developed it for Phonon and says it supports more than 10 backends, including execution on CPUs, GPUs, and NPUs. Listed options include WebAssembly, CUDA, Vulkan, Metal, WebGPU, Apple Core ML, and Snapdragon.
Use Transformers if you need a PyTorch implementation, batch generation on GPUs, or integration with existing model workflows. Phonon support is documented from Transformers v5.21 onward, including support for full-graph torch.compile. Check that your installed version includes the required classes.
For an initial integration, load the checkpoint and processor, load a precomputed voice file, prepare the input text, and generate audio latents. The processor then decodes the latents into playable audio. The supplied voice representations are tensors, so do not assume that loading an arbitrary recording directly will create a new voice.
Check Resource Usage on Real Devices#
The published package sizes help with download planning. They do not indicate peak memory usage. In the provided Transformers example, the English generation model has 30,788,354 parameters, and the audio tokenizer has another 10,305,313.
Measure the entire app on the lowest-spec hardware you support. Record time to first audio, sustained generation speed, peak memory usage, and power consumption. A short demo on a powerful laptop is not enough to establish adequate performance on a smartphone or single-board computer.
Offline operation also needs to be verified across the entire app. Download the required model and voice data first, disable the network, and run the full speech pipeline. Phonon can synthesize speech locally, but where users’ text is sent also depends on your app’s telemetry and surrounding services.
Test with User Input#
According to Gradium, Phonon can handle phone numbers, dates, and codes. This is useful, but reading conventions vary by language and product. Test the actual formats your interface accepts.
Create a small evaluation set containing everyday sentences, specialized terminology, formatted numbers, and long passages. Listen for missing words, inappropriate pauses, and inconsistent pronunciation. Process the same inputs with your current speech system to make a comparison relevant to your product’s needs.
Next Steps#
Run Phonon on your lowest-spec supported device using 20 representative sentences that include dates and product terminology, and record pronunciation errors and time to first audio before deciding whether to integrate it.
Source: https://huggingface.co/blog/phonon-tts
Contents
