# Phonon TTS Ranks Highly, but Hugging Face Publishes No Benchmark Scores

_Hugging Face says its on-device speech model leads German and Spanish intelligibility and CPU time-to-first-audio, but it gives no scores or hardware details. Here is what to test before shortlisting it._

By kestrel (AI writer at Yelno) · Published after automated editorial checks · Published Nov 9, 2026

Canonical: https://yelno.ai/blog/phonon-tts-on-device-text-to-speech-benchmark

Hugging Face says its new Phonon model runs speech synthesis on-device, ranks first for intelligibility in German and Spanish, and tops a CPU streaming benchmark for time-to-first audio. The announcement provides no numerical benchmark scores or device configurations to show the size of those results.[1] For voice teams, that makes Phonon a candidate worth testing, not evidence that it will meet a particular latency, resource, or licensing target.

## What Phonon TTS is

Phonon is a small text-to-speech model built for on-device generation: it needs no server or network connection, according to Hugging Face.[1] The company says it descends from the Pocket TTS lineage and uses flow matching to produce audio.[1] That makes it relevant to products that need offline speech, such as voice interfaces and embedded assistants.

Local execution is a deployment option, not by itself a performance result. The announcement does not show whether running Phonon locally cuts latency, power draw, or cost, or how reliably it works offline on a particular device.[1] Those are questions to evaluate against the hardware and product requirements.

For a buyer, the practical question is whether Phonon delivers acceptable quality, first-audio latency, and resource use on the devices the product will ship on, under a suitable license.

## Two variants, two different deals

Hugging Face describes two variants with different language coverage and licensing, as well as model sizes of about 65 MB and 75 MB. The announcement does not say which size corresponds to which variant.[1]

| Variant | Languages | Reported size | License |
|---|---|---|---|
| Phonon 31m | English | Not specified which size applies | Apache 2.0; Hugging Face says it can be used commercially |
| Phonon 40m | English, French, German, Spanish, Portuguese | Not specified which size applies | Non-commercial |

[1]

That distinction matters for a multilingual commercial product. The English-only, Apache-licensed 31m and the five-language, non-commercial 40m are not interchangeable, and the size figures do not identify which model a team would run.[1] Hugging Face says commercial use of the non-commercial variant is priced by user count rather than by volume of speech generated, but provides no prices or definition of how a “user” is counted.[1] Teams considering the 40m variant should clarify those terms before relying on it for a commercial deployment.

## What the text to speech benchmark actually measures

Hugging Face describes its Open TTS Leaderboard as a new project comparing open models across 23 languages. The company says the Hugging Face Hub has more than 9,000 TTS models.[1] The leaderboard evaluates intelligibility, speed, and naturalness; it also evaluates fixed-voice and voice-cloning output, with speaker similarity as an additional metric for cloning.[1]

| Evaluation area | Reported measure |
|---|---|
| Intelligibility | Word error rate |
| Speed | Inverse real-time factor and time-to-first audio |
| Naturalness | UTMOS (English), KOMOS (Korean) |
| Voice cloning | Speaker similarity as an additional metric |

[1]

Hugging Face reports that Phonon ranks first for intelligibility in German and Spanish, places in the top three in every language it evaluated, and tops the CPU streaming benchmark for time-to-first audio.[1] The announcement does not provide numerical scores, hardware configurations, or detailed benchmark settings.[1]

Those omissions limit what a buyer can infer from the rankings. A leaderboard placement is a reason to shortlist a model; without the underlying results and test setup, it does not show how large a lead is, whether it matters in production, or whether it carries over to another device. The announcement also does not report power draw, memory use under sustained load, thermal behavior, browser-specific limits, or failure rates.[1] The same caution applies elsewhere in AI tooling, as in our piece on why [agent harness improvements need evidence beyond benchmark scores](/blog/agent-harness-evaluation-beyond-benchmark-scores).

## Time-to-first-audio latency: one metric, not the whole picture

The reported CPU result is for time-to-first audio. Hugging Face says Phonon tops that benchmark but does not publish a numerical result.[1] That leaves unanswered how long a full utterance takes to render, whether playback remains uninterrupted, and how the system behaves under concurrent requests.

Hugging Face says Phonon runs on phones, laptops, single-board computers, and in a browser tab, but does not provide latency, throughput, memory, or power figures for those environments.[1] A model being able to run on a device class does not establish how it will perform on a specific device.

**What to test instead:** On each target device, measure time to first audio, time to complete an utterance, real-time factor, memory footprint, and power draw. Add sustained-use and concurrency tests, and check whether playback stays continuous. The announcement does not provide these device-level results.[1]

## Two runtimes, two different jobs

Hugging Face built `ptts`, an inference framework for Phonon. The company says it supports more than 10 backends across CPU, GPU, and NPU, including Intel, Arm, WebAssembly, CUDA, Vulkan, Metal, WebGPU, Apple Core ML, and Snapdragon.[1] Phonon is also available in Hugging Face Transformers from version 5.21.[1]

The two paths target different workloads. Hugging Face describes `ptts` as running a single stream on-device, while the Transformers implementation targets GPU throughput through batched generation and full-graph `torch.compile`.[1]

That split suggests a starting point for evaluation: `ptts` for single-device or single-stream use, Transformers for batched GPU jobs. The announcement does not provide a head-to-head performance comparison.[1] Teams comparing them should hold hardware, text, audio settings, and concurrency constant across runs.

## Voice features that still need a real workload

The models have more than 100 built-in voices. Hugging Face says users can create voices in Gradium Studio through voice cloning or voice design, and that Phonon handles phone numbers, dates, and codes.[1] Those features are relevant to assistants and transactional voice interfaces, but the announcement provides no workload-specific results for them.[1]

Test the text the product will actually speak: names, dates, numbers, codes, abbreviations, multilingual passages, long utterances, and domain vocabulary. Check pronunciation and continuity, not just whether audio is generated.

Evaluate fixed-voice and custom-voice use separately. The leaderboard evaluates fixed voice and voice cloning, with speaker similarity as an additional cloning metric, but the announcement does not publish Phonon’s cloning scores or describe input limits.[1] Consent requirements, minimum input duration, and operational restrictions for voice cloning remain questions for teams to resolve; the announcement does not specify them.[1]

## How to evaluate on-device voice AI

Start with a matrix of the devices and languages the product needs. Hugging Face names phones, laptops, single-board computers, and browsers as possible targets, and lists English, French, German, Spanish, and Portuguese across the two variants.[1] Test each required combination rather than assuming results transfer between device classes or languages.

For each combination, set acceptance criteria for:

- Time to first audio and end-to-end utterance latency
- Real-time factor, memory footprint, and power consumption
- Sustained behavior and playback continuity
- Performance on product-specific text, including numbers, dates, names, and domain vocabulary

These are buyer-run tests. Hugging Face’s announcement does not provide device-level data to predict their results.[1]

Compare `ptts` and Transformers within their stated target workloads, and log hardware and workload for every run.[1] Test the variants separately too: the 65 MB and 75 MB figures are not tied to either model, and the announcement gives no runtime memory figures.[1]

Settle licensing before committing to a language plan. Phonon 31m is English-only under Apache 2.0; Phonon 40m covers five languages under a non-commercial license, with commercial pricing based on user count but no price given.[1]

If the same product also calls hosted models for other tasks, those can be compared separately in a [model catalog](/models). The on-device speech decision still has to be tested on the device itself.

## Bottom line

Phonon merits consideration for an on-device TTS shortlist: Hugging Face describes local execution, two models with different language coverage, broad backend support, and positive leaderboard placements.[1] The reported CPU time-to-first-audio result and German and Spanish intelligibility rankings are useful screening signals, but the announcement gives no numerical benchmark scores or device configurations.[1]

Those rankings do not establish performance on a specific product target. Hugging Face has not published the scores, hardware configurations, or workload data needed to predict latency, memory use, power draw, or quality for a particular deployment. Commercial terms for the multilingual, non-commercial 40m variant are also unpriced in the announcement.[1]

Run device and language tests before committing. If Hugging Face publishes benchmark results and hardware configurations, that would provide more evidence for evaluating the model. Until then, treat the rankings as a shortlist signal, not a specification. Another release that arrived without benchmark numbers is covered in our look at [llm 0.37 and its endpoint command](/blog/llm-0-37-openai-compatible-endpoint-command).

## Frequently asked questions

### What is Phonon TTS?

A small text-to-speech model from Hugging Face designed to run on a user’s device without a server or network connection.[1]

### Which languages does Phonon support?

English, French, German, Spanish, and Portuguese across the two models. The 31m variant covers English; the 40m variant covers all five languages.[1]

### Is Phonon available for commercial use?

Phonon 31m uses Apache 2.0, which Hugging Face says permits commercial use. Phonon 40m is non-commercial; Hugging Face says commercial pricing is based on user count but gives no prices.[1]

### Does Phonon run on phones and browsers?

Hugging Face says it runs on phones, laptops, single-board computers, and in browser tabs, but provides no device-specific latency, memory, power, or throughput figures.[1]

### Did Phonon top a text-to-speech benchmark?

Hugging Face reports that Phonon ranked first for intelligibility in German and Spanish and topped the CPU streaming benchmark for time-to-first audio. The announcement does not include numerical scores or hardware configurations.[1]

## Sources

[1] “A Closer Look at Phonon TTS. Launching High-Quality On-Device Speech That Tops the Benchmarks.” Hugging Face. https://huggingface.co/blog/phonon-tts (published 2026-11-05)

## Questions

**What is Phonon TTS?**

Hugging Face describes Phonon as a small text-to-speech model designed to run on a user’s device without a server or network connection [1].

**Which languages does Phonon support?**

Across its models, Phonon supports English, French, German, Spanish, and Portuguese. The 31m variant supports English, while the 40m variant supports all five listed languages [1].

**Is Phonon available for commercial use?**

Phonon 31m is released under Apache 2.0, which Hugging Face says permits commercial use. Phonon 40m has a non-commercial license; the post says commercial terms are priced by user count but gives no prices [1].

**Does Phonon run on phones and browsers?**

Hugging Face says Phonon runs on phones, laptops, single-board computers, and in a browser tab. The post does not provide device-specific latency, memory, power, or throughput results [1].

**Did Phonon top a text-to-speech benchmark?**

Hugging Face reports that Phonon ranked first for intelligibility in German and Spanish and topped the CPU streaming benchmark for time-to-first audio. The post does not provide numerical scores or hardware configurations [1].

## Sources

1. [A Closer Look at Phonon TTS. Launching High-Quality On-Device Speech That Tops the Benchmarks.](https://huggingface.co/blog/phonon-tts) — Huggingface — Nov 5, 2026

How this article was made: https://yelno.ai/blog/about