Recent full-stream text-to-speech (TTS) models enable low-latency personalized synthesis as input text arrives. However, sustaining generation beyond the lengths seen during training without additional training remains largely unexplored. We systematically analyze VoXtream2, which extends generation by periodically discarding past context while retaining the acoustic prompt. We find that this strategy does not prevent degradation: pronunciation accuracy deteriorates near the end of the training range due to the range of positions in which the model operates rather than the loss of past context, while speaker similarity gradually declines throughout generation. Based on these findings, we propose two training-free modifications that reduce the attention window and strengthen attention to the acoustic prompt. Experiments show that these modifications preserve pronunciation accuracy and speaker similarity throughout long-form generation while retaining the naturalness of the original model. Ablations further show complementary effects of the two modifications.
Figures & tables
Figure 1: Segment-wise WER and SPK-SIM of VoXtream2, together with the corresponding ground truth speech.
Figure 2: Expected number of phonemes advanced per frame versus position for W=400 , 625 , and 900 . Shaded bands show standard errors over 8 randomly selected utterances.
Model
Metric
Objective Metrics
Subjective
Overall
Segment (s)
MOS ↑
0–30
30–60
60–90
90–120
120–150
150–180
180–210
Ground Truth (GT)
WER ↓
1.86
1.73
1.62
1.93
1.99
1.78
1.66
2.18
4.39±0.05
SPK-SIM ↑
0.832
0.849
0.837
0.834
0.834
0.831
0.832
0.826
[0.3pt/1.5pt] GT + Mimi Codec [ 2 ]
WER ↓
1.95
1.80
1.60
1.73
2.03
2.49
1.75
1.89
4.10±0.07
SPK-SIM ↑
0.754
0.771
0.760
0.756
0.757
0.753
0.754
0.748
Table 1: Overall and segment-wise performance on long-form speech synthesis. Lower WER (%) and higher SPK-SIM and MOS indicate better performance. MOS is reported as mean ± 95% confidence interval on a 5-point scale.
Figure 3: SPK-SIM over time with W=400 when the acoustic prompt is attended and when it is masked after 30 s. Error bars show standard errors over 8 randomly selected utterances.
Setting
Value
WER (%) ↓
SPK-SIM ↑
Window size (No bias)
Default
W=625
8.63
0.659
[0.3pt/1.5pt] Window
W=500
5.42
0.668
W=375
4.82
0.683
W=250
6.51
0.703
Attention bias ( W=375 )
Table 2: Ablation study on window size and attention bias. W , b , and ψtarget denote the window size, bias strength, and target attention mass, respectively.
Neural Text-to-Speech (TTS) systems achieve remarkable quality on short utterances but long-form speech generation shows prosodic drift, speaker inconsistencies and sentence boundary artifacts. Existing approaches either compress sequences, increase context length or naively concatenate independently synthesized chunks. We present an inference-time approach called MagpieTTS-LF that enables MagpieTTS to produce coherent long-form speech without model retraining. Our method introduces three key innovations: (1) soft attention priors to guide monotonic alignment while preserving past and future context; (2) a stateful inference algorithm that maintains context across sentence chunks, ensuring prosodic continuity; (3) history-aware text encoding that uses past text for discourse-level prosodic planning. Experiments on long texts show significant improvements in long-range intelligibility, prosodic coherence, speaker consistency, and boundary naturalness compared to other baselines.
Long-form text-to-speech (TTS) enables multi-turn conversations with consistent prosody and higher quality voice cloning from longer reference audio. Recent open-weights autoregressive TTS models such as Qwen3-TTS and VoxCPM2 attain state-of-the-art word error rate (WER) and speaker similarity (SIM) on short-form prompts but significantly deteriorate when used with long-form prompts. We propose Localized Attention-Constrained Inference (LACI), an inference-only method to detect TTS errors in near real-time, roll back to the error onset and regenerate with temporary guardrails, adding negligible computational overhead. Using LACI, we improve worst-of-N WER across 10 RNG seeds for Qwen3-TTS-0.6B from 35.2% to 3.4% on prompts longer than 1500 words, even surpassing its short-form reliability of 5.4% on prompts with fewer than 500 words. To demonstrate the efficacy of LACI on voice cloning reliability, we propose a sliding-window version of the SIM metric that we call wSIM. wSIM exposes several novel failure patterns that are not captured by SIM. LACI improves worst-of-N wSIM from 0.01 to 0.47 on 120 seconds of reference audio while reducing the rate of catastrophic generations with WER above 30% from 26% to below 1%
Rongxiang Wang, Berkin Durmus, Aysegul Orhon +2
Argmax, Inc. · University of Virginia · Bilkent University
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.
Nikita Torgashov, Gustav Eje Henter, Gabriel Skantze
Department of Speech, Music and Hearing, KTH Royal Institute of Technology, Sweden