ZONOS3-Mini is an ultra-compact, ultra-fast, zero-shot voice-cloning TTS model. The model can run in real time on a laptop or smartphone. For its latency and size, it achieves state-of-the-art accuracy and cloning fidelity. To achieve this we propose a latent MoE backbone with only 886M parameters (100M active per token) and a new audio codec, Zyphra Audio Codec 1 (ZAC-1), which represents 44.1 kHz audio at a 15 Hz frame rate. On the new ZTTS2-Eval benchmark, ZONOS3-Mini reaches a 6.71% macro-average WER over 55 language sets (35 clean and 20 in the wild), which is competitive with larger and slower state-of-the-art models. We release the model weights, inference, and training code.
Figures & tables
Figure 1: Quality versus speed on one H100 80GB. Left: time to first audio (p50) at 8 concurrent requests; right: peak throughput at each model’s best-throughput concurrency. Quality is 21 z(macro speaker similarity) −21 z(macro WER) over the 55 ZTTS2-Eval language sets; the dashed line is the Pareto frontier.
Figure 2: ZAC-1. One semantic and nine residual codes per 15 Hz frame are passed to the backbone.
WER ↓
Sim ↑
Codec
Frames/s
clean
wild
clean
wild
Fish S1–DAC
21.5
4.30
5.28
84.8
79.0
ZAC-1
15.0
4.06
5.54
83.8
76.1
MOSS, 14 RVQ
12.5
5.06
5.67
84.6
80.3
DAC
86.1
42.2
51.7
78.8
63.7
Table 1: Downstream codec comparison: 50k-step TTS models on the four-language ZTTS2-Eval sets. Best in bold.
Table 2: Codec reconstruction quality on ZTTS2-Eval clips and DAPS clean. Mel is the log-mel L1 distance. Best in bold, second best underlined.
Figure 3: Learning rate, training loss and normalized router entropy over pre-training, fine-tuning, the speaking-rate stage and post-training (DPO and GRPO). The fine-tune loss starts higher due to the addition of speaker-projection and control-row tables and a change to the data mix. Router entropy is the mean and range over the ten MoE layers.
Clean (35 languages)
In the wild (20 languages)
Macro
Model
Conditioning
WER ↓
Sim ↑
WER ↓
Sim ↑
WER ↓
Sim ↑
ZONOS3-Mini
embedding
8.28
83.1
7.52
74.0
8.00
79.8
ZONOS3-Mini
raw audio
7.95
77.7
12.57
55.1
9.63
69.5
ZONOS3-Mini + PT
embedding
7.79
83.4
4.81
77.3
6.71
81.2
ZONOS3-Mini + PT
raw audio
7.39
76.5
5.48
58.1
6.70
69.8
Table 3: ZTTS2-Eval results at CFG 1 over all 55 parts. Best in bold, second best underlined.
Params
Clean
In the wild
Macro
System
total / active
WER ↓
Sim ↑
WER ↓
Sim ↑
WER ↓
Sim ↑
Ground truth
–
9.66
70.7
9.19
79.3
9.49
73.8
With reference transcript
Qwen3-TTS 1.7B [ 16 ]
1.7B
22.80
80.6
14.61
74.5
19.82
78.4
Fish S2 Pro [ 24 ]
4.4B
11.03
81.2
5.98
70.3
9.20
77.2
VoxCPM2 [ 58 ]
2B
12.47
86.2
5.12
77.8
9.80
83.1
Table 4: Comparison with external systems on all ZTTS2-Eval languages (35 clean and 20 in-the-wild parts). Systems are grouped by whether they are conditioned on the transcript of the reference clip. Best in bold, second best underlined, within each group. Per-language results for every system are given in Appendix 14 .
Clean, 4 languages
In the wild
System
Voice
WER ↓
Sim ↑
WER ↓
Sim ↑
Cartesia Sonic 3.6
clone
2.49
85.0
3.30
76.4
Inworld TTS 2
clone
2.42
71.5
3.89
61.4
ElevenLabs v4
clone
2.07
82.9
2.97
–
Gemini 3.8 Flash
preset
2.55
–
4.67
–
\zyoursrow ZONOS3-Mini + PT, embedding (ours)
clone
2.55
83.1
5.00
77.5
Table 5: Comparison with closed-source systems on ZTTS2-Eval Mini, a subset of ZTTS2-Eval that reduces evaluation cost, restricted to the items every system generated. Best in bold, second best underlined.
In the wild
Clean
Arm (cumulative)
WER ↓
Sim ↑
UTMOS ↑
WER ↓
Sim ↑
UTMOS ↑
Original (ZONOS2 block)
64.6
45.1
2.10
39.1
59.2
2.86
Latent MoE
83.2
44.2
1.95
33.7
69.3
2.87
+ linear router, no EDA
43.5
54.5
2.19
26.9
65.8
2.99
+ quantile balancing
39.0
56.3
2.20
33.7
60.0
2.93
+ XSA
49.1
53.1
2.19
31.2
67.0
2.95
Table 6: 150k backbone ablation on the four-language ZTTS2-Eval sets. Best in bold.
Configuration
Family
Frames/s
Codebooks
Codebook sizes
kbit/s
WER ↓
Sim ↑
Fish S1 codec [ 24 ]
S1– DAC
21.53
10
1×4096+9×1024
2.20
4.79
81.87
ZAC-1
dual hybrid
15.00
10
10×4096
1.80
4.80
79.94
Dual hybrid, 10 s
dual hybrid
15.00
10
10×4096
1.80
5.81
74.54
Dual hybrid R1024
dual hybrid
15.00
10
1×4096+9×1024
1.53
6.53
67.83
Hybrid, 15 Hz, 10 CB
hybrid
15.00
10
10×4096
1.80
7.25
68.37
Hybrid, 12.5 Hz, 10 CB
hybrid
12.50
10
10×4096
1.50
7.22
65.11
Table 7: Downstream codec ablation.
Figure 4: SuperBPE vocabulary sweep: matched 50k-step runs on the S1–DAC codec scored on the ZTTS2-Eval clean and in-the-wild sets (four-language macro over en, zh, es and ja; bars are 95 % confidence intervals of the clipped WER ).
Pre-train 375k
Fine-tune 60k
Rate stage 20k
Nodes × GPUs
14 × 8 H100
14 × 8 H100
14 × 8 H100
Positions / GPU / step
106,496
106,496
106,496
Positions / step
11.93M
11.93M
11.93M
Steps
375,000
60,000
20,000
Muon / AdamW / bias LR
0.02 / 0.002 / 0.002
0.0035 / 0.00035 / 0.00035
0.0035 / 0.00035 / 0.00035
Warm-up / schedule
250 / cosine to 10−6
100 / cosine to 0
100 / cosine to 0
Table 9: Training hyper-parameters of the released checkpoints.
Model
Metric
en
zh
de
es
fr
it
ja
ko
ru
pt
ar
hi
tr
pl
id
th
fil
nl
Reference
Ground truth
WER ↓
4.30
3.02
4.64
2.89
4.08
4.05
4.16
3.20
5.67
4.18
13.69
14.61
9.60
4.48
7.31
5.44
13.43
6.83
Spk. sim. ↑
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
With reference transcript
Qwen3-TTS 1.7B [ 16 ]
WER ↓
2.37
4.64
3.17
2.26
3.07
2.36
3.71
4.32
5.21
3.18
31.66
29.32
37.49
51.17
8.18
5.64
16.18
7.39
Spk. sim. ↑
80.0
85.3
80.4
84.6
81.6
82.3
86.8
85.3
82.3
83.8
82.4
75.5
80.1
76.5
84.1
86.9
87.6
80.9
Table 10: ZTTS2-Eval clean set (FLEURS-R) zero-shot results per language, part one of two: clipped WER (%, CER for zh and ja) and ReDimNet speaker similarity (×100). Systems are grouped by whether the harness gives them the reference transcript, as in Table 4 . Best cloning system per language and metric in bold, second best underlined; the ground truth is shown for context and excluded from the ranking. A dash marks a language the system does not support.
Model
Metric
vi
cs
sv
da
fi
el
hu
ro
uk
nb
he
bg
hr
sk
ca
ms
fa
Reference
Ground truth
WER ↓
4.35
12.31
7.99
11.71
12.87
16.69
18.73
6.85
6.31
11.70
28.97
14.70
13.35
12.20
7.26
8.15
28.50
Spk. sim. ↑
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
–
With reference transcript
Qwen3-TTS 1.7B [ 16 ]
WER ↓
26.48
40.73
16.15
29.93
37.75
76.66
49.35
24.83
13.83
20.80
91.64
29.28
24.10
34.40
8.91
8.05
43.65
Spk. sim. ↑
83.5
74.9
79.4
74.4
81.3
68.2
76.9
82.9
83.0
77.7
71.1
77.0
81.9
77.7
80.3
85.7
80.2
Table 11: ZTTS2-Eval clean set (FLEURS-R) zero-shot results per language, part two of two: clipped WER (%, CER for zh and ja) and ReDimNet speaker similarity (×100). Systems are grouped by whether the harness gives them the reference transcript, as in Table 4 . Best cloning system per language and metric in bold, second best underlined; the ground truth is shown for context and excluded from the ranking. A dash marks a language the system does not support.
Model
Metric
en
zh
de
es
fr
it
ja
ko
ru
pt
ar
hi
id
tr
tl
pl
th
nl
uk
vi
Reference
Ground truth
WER ↓
3.82
8.78
6.25
5.08
5.57
6.25
6.46
9.02
9.53
5.58
11.62
13.34
15.18
17.24
12.54
9.90
8.74
6.51
12.20
10.28
Spk. sim. ↑
76.3
80.6
77.9
78.9
77.7
79.1
80.4
79.0
79.2
80.1
79.8
78.0
78.8
80.2
78.3
80.2
81.5
76.1
80.8
82.2
With reference transcript
Qwen3-TTS 1.7B [ 16 ]
WER ↓
2.63
5.79
2.92
2.25
2.69
2.99
4.46
4.78
5.05
4.23
24.76
67.34
11.17
35.52
10.69
49.40
7.19
7.38
13.26
27.78
Spk. sim. ↑
74.7
79.4
76.2
76.4
75.5
76.5
78.7
78.8
78.7
76.3
75.3
45.5
76.3
73.5
73.5
70.4
79.3
73.5
77.7
73.6
Table 12: ZTTS2-Eval in-the-wild set (VoxBlink2) zero-shot results per language: clipped WER (%, CER for zh and ja) and ReDimNet similarity to the prompt (×100). Grouping and marking as in Table 10 .
We present ZONOS2 8B, our latest TTS model, which achieves state-of-the-art naturalness, prosody, and voice cloning fidelity. We improve upon Zonos-v0.1 across scale, data, and training recipe. We scale the model from 1.6B to 8B total parameters (900M active) with a novel mixture-of-experts (MoE) backbone, improving inference latency and throughput. We expand our training corpus from 200K to over 6M hours using a new data processing pipeline, and we simplify our post-training and conditioning recipes to improve naturalness and voice cloning fidelity. We evaluate ZONOS2 8B on quality, speaker similarity, WER, and ZTTS1-Eval, our novel TTS benchmark, where it performs competitively with state-of-the-art systems while maintaining good streaming latency. We release our model weights and example inference code under an Apache 2.0 license on GitHub and Hugging Face.
We present M5-TTS, a compact text-to-speech system supporting multilingual and Chinese multidialect synthesis, zero-shot voice cloning, speaker-specific multilingual generation, and CPU streaming. Its 50M-parameter language model, excluding the speech codec, combines a temporal backbone with a local multi-token prediction module over 16 codebooks. GRPO post-training combines global speaker-similarity and pause feedback with a token-level CER process advantage. Character-level error comparisons provide local feedback, with negative credit propagated backward within a clause and projected onto audio-token positions. M5-TTS achieves 1.37% CER on Seed-ZH and 2.91% WER on Seed-EN, corresponding to relative error reductions of 62.2% and 35.3% over MOSS-TTS-Nano with half the language-model parameters. It also achieves lower content error rates in five of six evaluated MiniMax languages, while difficult Chinese inputs remain a limitation. On four Intel Xeon Gold 6530 cores, it achieves a real-time factor of 0.653 and a warm first-chunk latency of 225ms. A 10-listener evaluation gives M5-TTS higher speaker-similarity and intelligibility ratings than MOSS-TTS-Nano, with a lower naturalness rating. We will release training, post-training, and inference code, together with voice-cloning and speaker-specific checkpoints, to support reproducible research on compact speech generation for on-device voice interaction.
Wenhao Li, Jingbin Hu, Shengming Ling +20
Shenzhen Pimei Technology Co., Ltd. · Moonstep AI · WeNet Open Source Community
TTS systems with autoregressive semantic modeling have demonstrated strong zero-shot voice cloning performance and rich expressive variation, but their sequential decoding incurs substantial latency. Non-autoregressive alternatives offer much faster generation, yet often rely on more restrictive reference conditioning, such as requiring transcripts of the reference speech during inference. We present Tacit-TTS, an efficient transcript-free zero-shot voice cloning system distilled from IndexTTS2. Our model replaces autoregressive text-to-semantic decoding with masked non-autoregressive generation, introduces training-free acoustic length estimation, and accelerates the flow-matching renderer through ReFlow distillation. Across two English and two Mandarin datasets, Tacit-TTS achieves competitive zero-shot quality while generating speech over 10x faster than IndexTTS2 for utterances longer than 5 seconds. Its transcript-free conditioning further supports cross-lingual and non-lexical references. We validate this capability using references from eight other languages, infant babble, and synthetic gibberish, where transcript-dependent systems often degrade or fail due to unreliable ASR transcripts.