Speaking Longer Than Learned: Training-Free Long-Form Generation for Full-Stream Text-to-Speech

1. INTRODUCTION

• Long-form synthesis beyond training lengths is underexplored, while paired speech data is scarce and concatenating segments can introduce speaker inconsistencies and boundary artifacts.The paper frames this as a length-extrapolation problem.Source: §1 · p. 1
• To the authors’ knowledge, training-free length extrapolation for pretrained full-stream TTS under continuous long-form generation had not been explored.Prior training-free work includes MagpieTTS-LF, while streaming approaches cited here require additional training for bounded-context operation.Source: §1 · p. 1
• The paper analyzes VoXtream2, finding that pronunciation degrades near the training-range end because of operating positions, while speaker similarity declines gradually despite cache resets.The proposed changes shrink the attention window and bias attention toward the acoustic prompt.Source: §1 · p. 1
• The proposed methods preserve pronunciation accuracy and speaker similarity in long-form streaming generation, with comparable naturalness and complementary effects from the two modifications.Speaker-similarity loss also occurs in open-source TTS models trained on long-form speech.Source: §1 · pp. 1–2

2.1. VoXtream2

• VoXtream2 conditions on reference speech encoded as Mimi codec tokens and uses phoneme, temporal, and depth transformers to generate speech.Its duration tokens encode phoneme advancement, and training uses random 50-second crops of 625 frames.Source: §2.1 · p. 2
• Resetting at W retains the acoustic prompt and recent generated frames, reindexes them from zero, and recomputes their key-value states.The default audio window is W = 625; resets retain P prompt frames and τ = floor(W/2) recent frames.Source: §2.1 · p. 2

2.2. Analysis of Long-Form Generation Degradation

• With cache resets, WER rises from 2.74% in the first 10-second window to 11.23% at 40–50 seconds, while SPK-SIM declines from 0.760 to 0.695 and continues falling.The analysis uses LibriSpeech-Long recordings up to four minutes, with opening speech as the acoustic prompt and remaining transcript as target text.Source: §2.2 · p. 2, Figure 1 · p. 2

2.2.1. Position, Not Cache Reset

• Near position 625, phoneme advancement rises sharply; WER increases from 7.3% to 49.8% across that position, but changes from 47.8% to 44.2% around the later cache reset.With W = 900, generated audio is 70% as long as ground truth and deletion errors rise from 0.7% at W = 625 to 16.6%.Source: §2.2.1 · pp. 2–3, Figure 2 · p. 3

2.2.2. Attending to the Prompt

• Reducing W slows the steady SPK-SIM decline, but does not remove it; masking the prompt after 30 seconds drops SPK-SIM from 0.73 to 0.23 in the next 10-second window.The test retains at least 200 generated frames, corresponding to 16 seconds of speech, indicating that the acoustic prompt remains important throughout generation.Source: §2.2.2 · p. 3, Figure 3 · p. 3
• The proposed prompt-attention intervention adds a constant bias to prompt-key logits across TT heads and layers without retraining or architectural changes.The analysis also compares adaptive SpotLight steering, which raises prompt-key logits when prompt attention mass falls below a target.Source: §2.2.2 · p. 3

3.1. Experimental Setups

• Experiments use pretrained VoXtream2 without additional training and follow the LibriSpeech-Long protocol, evaluating WER, SPK-SIM, and MOS.The MOS evaluation used 158 listeners and yielded 325 ratings per model.Source: §3.1 · p. 3

3.2.1. Model Comparison

• WER falls from 8.63% to 3.58% and SPK-SIM rises from 0.659 to 0.759 with the method; MOS remains 4.03, matching VoXtream2.Over long generation, the method keeps WER below 4.0% and SPK-SIM nearly stable across 30s segments.Source: §3.2.1 · pp. 3–4, Table 1 · p. 4
• Qwen3-TTS SPK-SIM drops from 0.780 to 0.560, while the method changes from 0.770 to 0.754 across later segments.MOSS-TTS also shows decreasing SPK-SIM toward later segments, indicating that this decline occurs in models trained with long-form data too.Source: §3.2.1 · p. 4

3.2.2. Ablation Studies

• W = 375 gives the best ablation WER of 4.82; with that window, b = 2.0 gives the highest SPK-SIM of 0.766, while b = 1.5 gives the lowest WER of 3.58.At W = 375, RTF rises from 0.415 at W = 625 to 0.433 and remains real-time; the experiments use W = 375 and b = 1.5.Source: §3.2.2 · p. 4, Table 2 · p. 4

• Shrinking the attention window and biasing attention toward the acoustic prompt preserves pronunciation accuracy and speaker similarity throughout long-form generation while retaining the original model’s naturalness.Source: §4 · p. 4

• The MOS evaluation used an online survey with participant consent and collected no sensitive personal information.Source: §5 · p. 5