Full-stream text-to-speech (TTS) for interactive systems must start speaking with minimal delay while remaining controllable as text arrives incrementally. We present VoXtream2, a zero-shot full-stream TTS model with dynamic speaking-rate control that can be updated mid-utterance on the fly. VoXtream2 combines a distribution matching mechanism over duration states with classifier-free guidance across conditioning signals to improve controllability and synthesis quality. Prompt-text masking enables textless audio prompting, removing the need for prompt transcription. Across standard zero-shot benchmarks and a dedicated speaking-rate test set, VoXtream2 achieves competitive objective and subjective results against public baselines despite a smaller model and less training data. In full-stream mode, it runs 4 times faster than real time with 74 ms first-packet latency on a consumer GPU.
Figures & tables
Model
Open
Voice
Full
Dynamic
Weight
Cloning
Stream
Control
Parler-TTS [ 8 ]
✓
Spark-TTS-Ctrl [ 9 ]
✓
ControlSpeech [ 10 ]
✓
MaskGCT [ 11 ]
✓
✓
VoiceStar [ 12 ]
✓
✓
Table 1: Comparison of VoXtream2 with other speaking rate control models. Spark-TTS-Ctrl stands for controllable version.
Figure 1: Overview of VoXtream2 architecture.
Model
#Data(h)
#Params
SEED test-en
LibriSpeech-PC test-clean
Naturalness ↑
WER (%) ↓
SPK-SIM ↑
UTMOS ↑
WER (%) ↓
SPK-SIM ↑
UTMOS ↑
μ± 95% CI
Human
-
-
1.50
0.734
3.53
1.95
0.695
4.10
56.7 ± 2.6
CosyVoice2
167k Multi.
618M
2.27
0.658
4.16
1.93
0.655
4.38
66.2 ± 2.5
VoiceStar
65k EN
840M
2.20
0.605
3.92
2.14
0.603
4.25
60.0 ± 2.6
MaskGCT
100k Multi.
1048M
2.03
0.712
3.56
2.58
0.691
3.92
57.6 ± 2.6
Spark-TTS
102k Multi.
507M
1.85
0.572
3.94
2.11
0.578
4.35
63.2 ± 2.5
Table 2: Evaluation results of zero-shot TTS models. Bold for the best result, underline for the second-best result. Naturalness is measured as MUSHRA scores (0–100). VoXtream2-B denotes baseline model without acoustic prompt enhancement.
Figure 2: Speaking rate control mechanism.
Model
Slow
Normal
Fast
WER (%) ↓
Baseline
2.80±5.64
2.10±4.32
4.21±8.10
Prompt mask
2.14±3.27
2.20±3.54
2.13±3.73
Speaking rate
Baseline
3.15±0.82
3.48±0.80
3.58±0.84
Prompt mask
2.76±0.72
3.09±0.74
3.37±0.84
Table 3: Evaluation of prompt text masking on different prompt speaking rates. Mean ± STD is reported.
Model
FPL(ms) ↓
RTF ↓
CosyVoice2: TRT
837
0.546
Kyutai-TTS: CG
277
0.272
VoXtream: TC
78
0.223
VoXtream2: CG
74
0.256
VoXtream2: TC
63
0.173
Table 4: Performance in mixed precision on an RTX3090 GPU in full-stream. TC denotes Torch Compile, CG denotes CUDA graphs, and TRT corresponds to TensorRT.
Figure 3: Evaluation of text chunk size in a full-stream.
Figure 4: Comparison of different TTS models across various speaking rates for utterance-level control.
Figure 6: Correlation between target and synthesized speaking rates across different TTS models.
Model
Corr.
WER (%) ↓
SPK-SIM ↑
UTMOS ↑
No SRC
-
3.26
0.624
3.86
S (1 SPS)
-
16.51
0.620
3.53
N (4 SPS)
-
3.20
0.626
3.83
F (7 SPS)
-
3.52
0.609
3.81
S → N → F
0.701
4.73
0.628
3.75
F → N → S
0.826
11.25
0.626
3.71
Table 5: Evaluation of dynamic speaking rate control. S, N, and F denote Slow, Normal, and Fast target speaking rates.
Figure 8: Examples of dynamic speaking rate control. S, N, and F denote Slow, Normal, and Fast target speaking rates.
Model
WER (%) ↓
SPK-SIM ↑
UTMOS ↑
VoXtream [ 31 ]
3.16
0.529
3.88
+ Data & updates
2.29
0.578
3.94
+ CFG text
1.37
0.578
3.91
+ CFG audio
1.62
0.661
3.65
+ CFG speaker
1.55
0.674
3.60
+ Speaker weight
1.45
0.679
3.56
Table 6: Modular analysis of VoXtream2 modifications.
TPS
LA
WER (%) ↓
FPL(ms) ↓
RTF ↓
-
∞
1.44
44
0.229
∞
3
2.29
75
0.256
40
3
2.29
75
0.254
20
3
2.98
67
0.259
10
1
56.13
56
0.262
10
2
22.89
56
0.274
Table 7: Effect of input text streaming rate (TPS) and phoneme look-ahead (LA) on full-stream TTS performance.
Figure 9: Influence of the CFG γ on the speaking rate control.