Smaller, clearer, better: Phonon Multilingual beats Magpie and NeuTTS Nano in French, German, and Spanish

Phonon, our 100M-parameter on-device Text-To-Speech model, produces up to 3.5x fewer word errors than NVIDIA's Magpie TTS in French, German, and Spanish, at 3.6x fewer parameters, and it leads the other on-device voice-cloning model, NeuTTS Nano, on both word error rate and speaker similarity. Phonon clones a reference voice across all five languages it supports: English, French, German, Spanish, and Portuguese.

Phonon is built on Continuous Audio Language Models [1] with flow-matching for waveform generation, first announced in April.

Our previous benchmark posts established Phonon as the leader on English. A model that tops the English charts can still mangle French liaison, German compound nouns, or Spanish stress patterns, so this post extends the evaluation to the other four languages: French, German, Spanish, and Portuguese, comparing Phonon against Magpie and NeuTTS Nano wherever a baseline exists.

Phonon at a glance

Spec	Phonon
Parameters	≈100M
Architecture	Continuous Audio Language Model with flow-matching for waveform generation
Text frontend	Standard text tokenizer, no per-language phonemizer
Languages	English, French, German, Spanish, Portuguese
Voice cloning	Yes, from a 10-second reference sample
Compute	Single CPU core
Memory footprint	~200 MB
Deployment	On-device and offline, no network round trip

How we evaluate multilingual Text-To-Speech quality

We use the same methodology as our previous Phonon evaluations, applied to one test set per language, and measure two axes.

Word error rate (WER) is computed by transcribing the generated speech back to text and comparing it against the original source with an edit distance from the jiwer package [2]. Lower is better. Speaker similarity is the cosine similarity between speaker embeddings extracted from the reference audio and from the generated speech. Higher is better.

For transcription we use whisper-large-v3 [3], which covers all evaluated languages. We deliberately avoid our own Speech-To-Text model here, so that modeling techniques shared between our TTS and STT models do not bias the results. For speaker similarity we use WavLM large [4]. All numbers were evaluated in July 2026.

Magpie TTS (NVIDIA) and NeuTTS Nano both cover French, German, and Spanish, and we evaluate each on all three. No baseline in our comparison set supports Portuguese, so we report Phonon's Portuguese numbers standalone. For English results, including the comparison against Kokoro and NeuTTS Nano, see our previous post.

One structural difference matters when reading the speaker similarity column. Phonon and NeuTTS Nano clone the reference speaker, conditioning on a reference sample for every utterance, so their similarity scores are directly comparable. Magpie synthesizes with a fixed voice and does not clone the reference, so speaker similarity does not apply to it; we mark those cells as such rather than report a misleading score.

French, German, Spanish, and Portuguese: Phonon vs Magpie and NeuTTS Nano

Phonon has the lowest word error rate of any baseline in French, German, and Spanish, and the highest speaker similarity than NeuTTS Nano. The table reports both metrics for each model on the languages it supports.

Language	Model	WER	Speaker similarity
French	Phonon	2.18%	78.98%
French	Magpie	3.48%	Voice cloning not available
French	NeuTTS Nano	4.11%	64.70%
German	Phonon	0.50%	73.25%
German	Magpie	1.47%	Voice cloning not available
German	NeuTTS Nano	1.95%*	61.89%
Spanish	Phonon	0.53%	81.50%
Spanish	Magpie	1.84%	Voice cloning not available
Spanish	NeuTTS Nano	2.33%	58.73%
Portuguese	Phonon	1.31%	85.38%

*NeuTTS Nano's raw German WER is 13.73%, inflated by a model hallucination on a single reference voice: NeuTTS Nano leaked that voice's prompt transcript into nearly every clip. Excluding the hallucinating voice, it averages 1.95% over the three remaining voices, which is the figure we report.

In German, Phonon's 0.50% WER is 2.9x lower than Magpie's 1.47%; in Spanish, 0.53% against 1.84%, a 3.5x gap; in French, the hardest set for both models, 2.18% against 3.48%. Magpie does not clone a reference voice, so there is no speaker-similarity comparison against it. NeuTTS Nano does clone, and Phonon leads it in every language on both axes: lower word error rate and higher speaker similarity (in Spanish, 0.53% vs 2.33% WER and 81.50% vs 58.73% similarity).

No baseline in our comparison set supports Portuguese, so Phonon's 1.31% WER and 85.38% speaker similarity, its highest of any language, stand without comparison for now. We will add baselines as competing models ship Portuguese support.

We attribute part of this gap to tokenization. Phonon uses a standard text tokenizer, while Magpie relies on a phonemizer. A phonemizer is a per-language component: its coverage and accuracy vary by language, and its errors propagate directly into synthesis. A text tokenizer transfers across languages with no language-specific frontend, which also makes Phonon more resilient to out-of-distribution text.

Parameter count vs WER: smaller model, more languages, lower error rate

Parameter count is the constraint that defines the on-device use case. A ~100M model fits comfortably in mobile memory and runs inference on a single CPU core. At 350M and above, memory and compute requirements start to dictate where the model can be deployed at all.

Model	Weights	Languages evaluated	Average WER
Phonon	≈100M	FR, DE, ES	1.07%
Magpie	357M	FR, DE, ES	2.26%
NeuTTS Nano	229M	FR, DE, ES	2.80%

On the three languages all three models cover, Phonon averages 1.07% WER against Magpie's 2.26%, roughly 2.1x lower, at 3.6x fewer parameters. NeuTTS Nano, at 229M (2.3x Phonon's size), averages 2.80% (its German figure excludes the hallucinating voice). Phonon also covers Portuguese (1.31% WER) with no baseline to compare against, plus English in our previous post. The takeaway is structural: Phonon covers more languages, at a smaller deployable size, with a lower error rate.

Why multilingual on-device TTS matters

On-device Text-To-Speech enables deployments that cloud APIs cannot serve: offline voice agents in vehicles and remote equipment, latency-sensitive products where a network round trip is unacceptable, and privacy-sensitive applications where audio cannot leave the device. None of these constraints stop at English-speaking markets. With a 100M-parameter model we can covers any of the five languages with one inference stack to maintain, and no per-language phonemizer to debug.

Try Phonon

To bring Phonon to your devices, get started at gradium.ai/on-device-tts, or reach the team at contact@gradium.ai. In the meantime, we continue to push for higher quality audio fitting the most compact model size.

Read the earlier posts in this series: Phonon update (May 2026) and Evaluating Phonon.

References

• [1] Kyutai, Continuous Audio Language Models, arXiv:2509.06926: arxiv.org/abs/2509.06926
• [2] jiwer, WER computation
• [3] whisper-large-v3, OpenAI
• [4] WavLM large, Microsoft