tts-bench — Scores

Objective scores over the same 5 prompts. UTMOS = predicted naturalness (higher better); WER = ASR word-error rate vs the intended text — a failure-detector, not a fine ranking (lower better); SIM = speaker similarity to the cloned reference (chris_hemsworth_15s, higher better). Switch Default / Cloning below; click any header to re-sort. Each score is the mean over the exact clips shown on Listen. Human votes are the preference ground truth; these objective metrics are backstops.

Default voice · naturalness + intelligibility

Model	Size	UTMOS ↑	WER ↓
Chatterbox	1.2B	4.423	0.061
Chatterbox Turbo	744M	4.274	0.074
Coqui XTTS-v2	750M	4.056	0.065
Dia 1.6B-0626	1.6B	2.387	0.317
dots.tts (soar)	2B	3.227	0.059
DramaBox	3.3B	4.304	0.093
F5-TTS v1	330M	4.081	0.107
Higgs Audio v3 TTS	4B	4.372	0.065
IndexTTS-2	1.5B	4.257	0.086
KittenTTS Nano 0.1	<100M	3.665	0.093
Kokoro	82M	4.302	0.065
LuxTTS	123M	3.540	0.092
Magpie-TTS	357M	4.199	0.087
Mars5-TTS	1.2B	3.540	0.361
Maya1	3B	4.487	0.066
MeloTTS	~52M	3.579	0.072
MiraTTS	0.5B	3.803	0.100
NeuTTS Air	748M	4.003	0.117
NeuTTS Nano	229M	3.572	0.066
OmniVoice	~1B	4.104	0.040
OuteTTS 1.0 1B	1B	4.386	0.070
Parler-TTS Mini v1	878M	3.757	0.149
Piper	~25MB	4.077	0.066
Pocket-TTS	100M	4.097	0.054
Qwen3-TTS 1.7B (CUDA-graph)	1.7B	4.323	0.065
Qwen3-TTS 1.7B Base	1.7B	4.276	0.096
Sesame CSM-1B	1B	4.152	0.114
Soprano 1.1 80M	80M	4.116	0.059
Step-Audio-EditX	3B	4.399	0.044
StyleTTS 2	~148M	4.259	0.158
Supertonic 3	99M	4.195	0.065
VibeVoice Realtime 0.5B	0.5B	4.043	0.148
VoxCPM2 2B	2B	3.480	0.023
Voxtral 4B TTS	4B	3.692	0.081

Cloning · fidelity + naturalness + intelligibility

Model	Size	SIM ↑	UTMOS ↑	WER ↓
Chatterbox	1.2B	0.627	4.278	0.073
Chatterbox Turbo	744M	0.666	4.043	0.096
Coqui XTTS-v2	750M	0.420	3.890	0.076
Dia 1.6B-0626	1.6B	0.551	3.580	0.343
dots.tts (soar)	2B	0.718	3.786	0.241
DramaBox	3.3B	0.778	3.889	0.080
Echo-TTS	2.8B	0.834	4.217	0.074
F5-TTS v1	330M	0.769	4.027	0.195
Fish Speech 1.5	~500M	0.623	3.673	0.109
Fish Speech S2-Pro	4B	0.725	4.270	0.058
Higgs Audio v3 TTS	4B	0.666	4.245	0.046
IndexTTS-2	1.5B	0.810	3.705	0.087
LuxTTS	123M	0.460	4.120	0.066
Mars5-TTS	1.2B	0.051	1.336	0.854
MetaVoice-1B	1.2B	0.382	3.796	0.337
MiraTTS	0.5B	0.505	3.726	0.181
MOSS-TTS v1.0	8B	0.731	4.059	0.030
MOSS-TTS v1.5	8B	0.746	4.118	0.046
MOSS-TTS-Nano	100M	0.642	3.950	0.147
NeuTTS Air	748M	0.298	3.510	0.569
NeuTTS Nano	229M	0.219	3.429	0.610
OmniVoice	~1B	0.751	4.042	0.125
OpenVoice v2	~100M	0.247	4.048	0.075
OuteTTS 1.0 1B	1B	0.466	3.598	0.145
Pocket-TTS	100M	0.549	4.075	0.067
Qwen3-TTS 1.7B (CUDA-graph)	1.7B	0.629	3.938	0.077
Sesame CSM-1B	1B	0.745	4.193	0.179
Step-Audio-EditX	3B	0.457	4.192	0.101
StyleTTS 2	~148M	0.297	4.328	0.115
VibeVoice 1.5B	1.5B	0.705	3.802	0.094
VibeVoice 7B	7B	0.671	3.965	0.051
VoxCPM2 2B	2B	0.533	3.596	0.040
ZipVoice 123M	123M	0.492	4.151	0.070
Zonos v0.1	1.6B	0.657	3.973	0.078

Scored over the 5 bench prompts (thin — WER is a failure-detector, not a fine ranking). Checkpoints: UTMOS utmos22_strong (SpeechMOS), SIM canonical UniSpeech-SAT wavlm_large_finetune, WER Whisper-large-v3. Method follows seed-tts-eval. Human votes are the preference ground truth; these are objective backstops.