The open-source TTS landscape now includes six distinct options: StyleTTS2, MATCHA-TTS, VITS, FastSpeech2, SoundStorm, and YourTTS. We compared them on a concrete test set, examining score-based diffusion, flow matching, GAN-based end-to-end synthesis, non-autoregressive FastSpeech-style models, codec language models, and multilingual zero-shot voice cloning. Their audio-generation methods, prosody assumptions, and typical failure points differ considerably.
We ran outputs from six public checkpoints through the Podonos API for a side-by-side view of each architecture's strengths and weaknesses. The results are Podonos MOS estimates from 50 Harvard Sentence List test sentences, all public domain. Every model used the same sentences and API version.
StyleTTS2-Base: naturalness strength, text-sensitive prosody
StyleTTS2 applies style diffusion within a style embedding space, giving it considerable room to create natural-sounding speech. It performed well for naturalness and pronunciation on our test set. Prosody varied more, especially in longer sentences with complex clause structures. Because prosody comes through style transfer from a reference or learned prior, familiar sentence forms work well, while unusual syntax can be less dependable.
Its main advantage is waveform quality and human-likeness, two difficult perceptual qualities to score highly. StyleTTS2 clearly improves on older architectures in these areas. For applications that prioritize naturalness and use controlled sentence complexity, it is a strong starting point.
MATCHA-TTS: steady prosody, moderate naturalness ceiling
MATCHA-TTS uses flow matching for acoustic modeling and delivers notably steady prosody. In the Podonos evaluation, its prosody sub-scores stayed consistent across the sentence types tested, with less variation than the diffusion-based approaches. Naturalness peaked below StyleTTS2, matching a common flow-matching trade-off: generation is faster and more controlled, but may not reach the upper naturalness levels of more expressive sampling.
MATCHA-TTS suits applications that value predictable prosody over maximum naturalness. Read-aloud products and other uses requiring consistent, non-variable output are good matches.
VITS-LJSpeech: excellent in-distribution results, clear off-distribution decline
VITS trained on LJSpeech scores very highly within its training distribution. LJSpeech is a single-speaker English audiobook recording, and VITS-LJSpeech was trained to reproduce that distribution closely. Our test sentences come from the Harvard Sentence List, which is also read-speech English, so the training overlap is substantial and the scores reflect it.
Its key limitation appears off distribution. Spontaneous speech patterns, casual register, and text with heavy punctuation variation score substantially below clean read-speech text. For applications using such inputs, strong standard benchmark results will not transfer. We also tested a small supplementary casual-register set and observed a meaningful naturalness drop compared with read speech. This reflects a training-distribution boundary, not an architectural flaw. Deployments using out-of-distribution text should be evaluated on representative production material rather than the standard test set.
FastSpeech2-EN: strong intelligibility, lower naturalness
FastSpeech2 is non-autoregressive and predicts duration and pitch explicitly. Those duration predictions support stable, correctly timed phoneme sequences, helping it score well for pronunciation and intelligibility. Naturalness trails the end-to-end architectures because explicit duration and pitch control can create more robotic prosody than learned end-to-end generation.
FastSpeech2 is useful when inference speed and control matter more than peak naturalness. Its explicit duration and pitch controls simplify adjustments to speaking rate and emphasis compared with black-box end-to-end models. Listeners can hear the naturalness gap versus StyleTTS2 or VITS, but that difference may be acceptable when fast, dependable output is the priority.
SoundStorm-ref: highest naturalness ceiling, needs reference audio
SoundStorm with reference audio conditioning achieved the highest naturalness scores in our set, showing what codec language models can do with fine acoustic detail from a reference. The required recording excludes applications without reference audio. It also uses more computation and runs more slowly than flow-matching or FastSpeech-family models.
SoundStorm-ref is strongest for voice cloning and voice consistency. When reference audio is available and a specific speaker must be reproduced with high naturalness fidelity, its naturalness ceiling was meaningfully above the other architectures tested. Prosody varied more, in part because codec language models derive it from both text and reference audio. That interaction makes generation more complex and gives the scoring model wider confidence intervals.
YourTTS-Multi: multilingual range with a naturalness trade-off
YourTTS supports multilingual, multi-speaker synthesis, and the benchmark shows the cost of that range. For English, its naturalness and prosody scores were below those of models optimized for one language. Sharing capacity across languages means the naturalness ceiling for any one language is lower than that of a single-language specialist.
That compromise can be worthwhile for multilingual teams. The relevant question is whether its naturalness cost fits your use case. For an application serving five languages, where listeners will not directly compare output with a single-language specialist, YourTTS-Multi is practical. For a high-quality single-language product that does not need multilingual support, paying that cost brings no benefit.
Lessons from the comparison
No architecture leads on every measure. The scores show clear trade-offs: end-to-end diffusion and flow-matching models raise naturalness while giving up some consistency and speed; codec language models offer the highest naturalness ceiling but need reference audio and more compute; non-autoregressive models provide dependable pronunciation and quick inference with lower naturalness; and multilingual models exchange depth of naturalness for language coverage.
Treat these scores as a starting point for your evaluation, not a final ranking. Our test set, the Harvard Sentence List, contains clean read-speech English. Different text distributions, registers, or domain vocabulary can produce different results. Before choosing whether to build or reuse, score the architectures under consideration on your own text distribution rather than relying on benchmark numbers from another test set.