- Rating
- 3
- Confidence
- 2
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper studies explicit duration control in end-to-end spoken dialogue models and introduces TiCo-Bench to evaluate this capability across four task categories. The proposed method, TiCo, inserts Spoken Time Markers (STMs) into intermediate text to represent cumulative speaking time. A first-stage supervised fine-tuning process learns these markers from self-generated speech with automatic timestamps, followed by GRPO training that rewards duration compliance and discourages malformed markers and repetition. The speech generator remains frozen, and reinforcement-learning rewards are computed from text rather than synthesized audio. Experiments on three backbones report approximately 2.7–3.5× reductions in duration error, broadly maintained automatic quality scores, and generalization across duration targets and instruction forms.
Critical strengths
-
Relevant problem and useful benchmark. The paper identifies a practical capability gap that existing SDMs handle poorly and provides a dedicated benchmark across multiple tasks. This establishes a useful evaluation resource beyond the proposed method.
-
Efficient method with consistent gains. STM-based rewards avoid expensive audio synthesis during reinforcement learning. Substantial improvements across three backbones, supported by ablations and generalization experiments, provide credible evidence of the method’s effectiveness.
Critical weaknesses
-
Insufficient validation of the STM-to-audio relationship. Rewards assess predicted timestamps rather than realized speech duration. Marker validity constraints do not ensure acoustic accuracy, and the reported marker–audio discrepancies warrant a clearer decomposition of planning error and realization error. Importantly, errors in generated markers should be distinguished from timestamp-label errors in the self-generated SFT data.
-
Incomplete evidence for quality preservation under duration constraints. Aggregate GPT-score and UTMOS do not establish that duration-compliant responses remain complete and natural, especially for bounded instructions evaluated only with duration metrics. Reported degenerate endings are not adequately quantified, and the absence of human evaluation and confidence intervals weakens quality-preservation claims and close baseline comparisons.
Questions for authors
-
Decomposition of duration error, with numbers rather than curves. The reward is defined against the final Spoken Time Marker, whereas the reported MAPE is computed against the realized audio duration. Figure 9 plots the error against the instructed duration and the error against the final STM, and the two curves nearly coincide, which suggests the planning error is small and the residual comes from the gap between planned and realized speech. However, this decomposition is only presented graphically. Could you report the two components numerically, namely the average error of the final STM relative to the instructed duration and the average error of the realized audio duration relative to the final STM? In addition, Table 8 reports a local alignment error of 2.65 s (Short) and 3.21 s (Long) aggregated over all markers. Could you report the same statistic for the final STM alone, since that is the quantity the reward actually verifies?
-
Possible circularity between training labels and evaluation of marker accuracy. Stage 1 constructs supervision using word-level timestamps from the Whisper medium, and Appendix F.1 evaluates marker accuracy against ASR word-level timestamps. To what extent does the reported marker accuracy reflect true acoustic timing, rather than agreement between the model and Whisper's own timing estimates? Could you cross-validate with an independent aligner, for example, the Montreal Forced Aligner already used elsewhere in the paper for segment-level prosodic statistics?
-
Quality metrics for the bounded instruction forms. Section 6.2 augments 100 audio queries with less than, more than, and in-between instructions and reports only MAPE and hinge MAPE, with hinge MAPE defined to be zero inside the admissible range. Could you additionally report the GPT score and UTMOS for these queries, in particular for the samples that fall inside the admissible range? Since hinge MAPE is identically zero there, the current metrics cannot establish whether an in-range response is complete and natural or degenerate.
-
Empirical rate of the degenerate endings. The Limitations section notes that, in order to end on time, the model sometimes closes with a generic offer of further help or an unfinished clause, and that the automatic metrics do not explicitly penalize this. What is the empirical rate of this behavior on TiCo-Bench, and is it correlated with how far the requested duration lies from the model's natural response length?
-
Statistical reliability of the main comparisons. Are the results in Table 1, Table 4, Table 7, and Table 12 obtained from a single run or averaged over multiple random seeds? If single-run, could you report confidence intervals or run-to-run variance, particularly for the comparison against the speaking-rate-controlled cascade, where the MAPE difference is about 1.5 points (16.19 vs. 17.67) while the UTMOS difference is 1.38 (4.04 vs. 2.66)? Note that the ±1 SEM shading in Figures 10 and 11 reflects sampling variability across evaluation samples and does not substitute for variance across training runs.
-
Justification of the reward hyperparameters. The copy-penalty tolerance is set to τ = 0.5 s, which is considerably smaller than the reported marker alignment error of 2.65 s to 3.21 s. Could you explain how τ and the Gaussian tolerance σ were selected, and whether performance is sensitive to these values?
6 comments
Response to Reviewer 2EfV (2/3): W2, Q3 and Q4
W2: Evidence for quality preservation under duration constraints
Thank you for raising this. We agree that aggregate scores alone do not show that the replies stay complete, so we measured the endings directly on all 2,000 replies per model.
Table R6-2EfV. Ending types on TiCo-Bench (Qwen2.5-Omni-7B, 2,000 replies per model).
| Model | Unfinished clause | Generic offer of help | Either |
|---|---|---|---|
| Qwen2.5-Omni-7B (base model) | 0.0% | 91.0% | 91.0% |
| Qwen2.5-Omni-7B + TiCo | 13.2% | 36.0% | 38.5% |
The generic offer is a habit of the base model, which ends 91% of its replies this way. TiCo reduces it to 36% and spends the time on the answer. The cost is the unfinished clause, which appears in 13.2% of TiCo's replies.
The unfinished ending can be largely removed. We retrained Qwen3-Omni with a system prompt that asks the model to end on a statement, without offering further help. The completion rate is the share of replies that end on a complete sentence.
Table R7-2EfV. Completion rate and duration error (TiCo-Bench, 2,000 prompts; MAPE of the first two rows from Table 1).
| Model | Completion rate | MAPE ↓ |
|---|---|---|
| Qwen2.5-Omni-7B + TiCo | 86.8% | 16.2% |
| Qwen3-Omni-30B + TiCo | 93.0% | 12.2% |
| Qwen3-Omni-30B + TiCo, trained to end on a statement | 99.2% | 17.7% |
The completion rate rises from 93.0% to 99.2%, and the duration error (17.7%) remains far below the base model's 42.1%. The unfinished ending is therefore not inherent to duration control: it can be trained away, at some cost in duration accuracy. We are repeating this on Qwen2.5-Omni.
Confidence intervals and human evaluation. With a paired bootstrap over queries, the overall GPT-score does not differ between the backbone and TiCo (3.31 vs. 3.32; difference −0.01, 95% CI [−0.06, 0.04]). Intervals for the other comparisons are in Q5. A human evaluation is in progress, and we will add it to the revision.
Q3: Quality metrics for the bounded instruction forms
Section 6.2 tests whether duration control generalizes to instruction forms never seen in training, and TiCo reduces hinge MAPE by 2.4–3.4× on them. We agree that a zero hinge MAPE says nothing about the quality of an in-range reply. We are now running GPT-score and UTMOS on these replies, for each bounded form and separately for the in-range replies, together with the ending analysis of Tables R6-2EfV and R7-2EfV, and will report them in the revised version.
In the meantime, the demo page provided in the paper includes audio samples under less than, more than and in-between instructions, where the replies remain complete and natural.
Q4: Empirical rate of the degenerate endings
The empirical rate is in Table R6-2EfV above: 13.2% of TiCo's replies end on an unfinished clause and 36.0% on a generic offer, against 0.0% and 91.0% for the base model.
Relation to the natural response length. We take the base model's duration on the same query as the natural length.
Table R8-2EfV. TiCo's ending types by the distance between the requested duration and the natural length.
| Requested minus natural length | Replies | Unfinished clause | Generic offer |
|---|---|---|---|
| More than 15 s shorter | 169 | 10.1% | 14.8% |
| 5 to 15 s shorter | 395 | 12.9% | 20.5% |
| Within 5 s | 488 | 18.4% | 37.9% |
| 5 to 15 s longer | 452 | 17.0% | 45.1% |
| More than 15 s longer | 496 | 6.2% | 45.4% |
The two endings behave differently. The generic offer depends on the direction: when the requested duration is shorter than the natural length, TiCo drops the offer (15–21%); when it is longer, there is room and the offer stays (45%). The unfinished clause does not grow with the distance: it is most frequent when the request is close to the natural length (17–18%) and least frequent at both extremes. Overall, the correlation between the absolute distance and a degenerate ending is −0.03.
Response to Reviewer 2EfV (3/3): Q5 and Q6
Q5: Statistical reliability of the main comparisons
All results in Tables 1, 4, 7 and 12 come from a single training run per model and a single generation pass, because RL post-training of 7B–30B SDMs is expensive. We add two measurements of uncertainty.
Sampling uncertainty and run-to-run variance. We report bootstrap 95% confidence intervals over the 2,000 queries. We also trained TiCo on Qwen2.5-Omni-7B a second time from scratch (new Stage 1 and Stage 2 runs).
Table R9-2EfV. Duration error with confidence intervals, and a second training run (TiCo-Bench, 2,000 prompts).
| Model | MAPE ↓ [95% CI] |
|---|---|
| Qwen2.5-Omni-7B (base model) | 43.7% [42.0, 45.5] |
| Qwen2.5-Omni-7B + TiCo, second training run | 17.6% [16.6, 18.6] |
| Qwen2.5-Omni-7B + TiCo, Table 1 | 16.2% |
The two training runs differ by 1.4 points, and the interval of one run is about ±1 point. The improvement over the base model is 26 points, more than ten times either source of variation. It is also replicated on three independently trained backbones with different architectures (Qwen2.5-Omni, Qwen3-Omni, Step-Audio-2-mini), each reducing MAPE by 2.7–3.5×.
Comparison with the speaking-rate-controlled cascade. We agree that the 1.5-point MAPE difference (16.19% vs. 17.67%) is within the run-to-run variation above, and we do not claim that TiCo is more accurate in duration than this cascade. The claim rests on naturalness.
Table R10-2EfV. Speaking-rate-controlled cascade and TiCo (from Tables 1 and 7).
| System | MAPE ↓ | UTMOS ↑ | ΔUTMOS |
|---|---|---|---|
| Cascade (Qwen), natural | 46.36% | 3.12 | – |
| Cascade (Qwen), adjusted | 17.67% | 2.66 | −0.46 |
| Qwen2.5-Omni-7B | 43.33% | 4.09 | – |
| + TiCo | 16.19% | 4.04 | −0.05 |
The absolute UTMOS (4.04 vs. 2.66) includes the contribution of each system's synthesizer: without any duration control the two systems already differ by 0.97. To isolate the cost of the control mechanism, we compare the change within each system. At comparable duration accuracy, rate adjustment lowers UTMOS by 0.46, whereas TiCo lowers it by 0.05. Both the 1.38 gap and the 0.46 drop are far larger than the sampling uncertainty of a mean over 2,000 utterances.
We will add confidence intervals to the tables and state the single-run setting in the revision.
Q6: Justification of the reward hyperparameters
τ and the alignment error measure different things. The copy penalty compares non-final markers with the instructed duration, not with the audio. It catches a textual shortcut: writing the target value early. τ therefore only needs to capture near-exact copies, and 0.5 s spans a few steps of the 0.1 s STM resolution. The 2.65–3.21 s alignment error is the distance between a marker and the audio, which the copy penalty does not use.
σ. σ = 5 s gives the main reward a useful slope across the 10–60 s range: the reward is 0.92 at a 2 s error, 0.61 at 5 s and 0.14 at 10 s.
Selection and sensitivity. Both are ordinary hyperparameters. We set them once and did not run a search, so we cannot report sensitivity. Table 12 does confirm the role of the copy penalty: the variants with it reach 15.38–19.07% MAPE, against 28.10–38.89% without it. We will state this in the revision.
Summary
We sincerely thank the reviewer for the time and patience in reading this long response, and for helping us make the manuscript more comprehensive. We will include these discussions (e.g. the decomposition of the duration error, the cross-validation with an independent aligner, the ending analysis and the confidence intervals) in the revised version. We are also running human evaluations.
Response to Reviewer 2EfV (1/3): W1, Q1 and Q2
We thank the reviewer for recognizing the practical value of TiCo and the relevance of the problem it addresses, as well as the consistent gains across three backbones, which the reviewer found to be credible evidence of the method's effectiveness. Below we address each concern.
W1: Validation of the STM-to-audio relationship
Thank you for raising this. W1 makes three points. Two are answered here; the decomposition into planning error and realization error is in Q1 below, and the independent aligner in Q2.
The reward is on the markers, not on the realized speech. We checked the realized audio at checkpoints along RL training.
Table R1-2EfV. Duration error and marker accuracy along RL training (retrained model, 320 prompts).
RL reduces the error of the realized audio from 44.0% to 15.4%, while the gap between the marker and the audio stays at about 12%. The realization error is already present after Stage 1, and optimizing the markers does not widen it.
Label error against generated-marker error. We compared Whisper, which produces the labels, with the Montreal Forced Aligner (MFA) on the Stage 1 training data.
Table R2-2EfV. Whisper against MFA on the Stage 1 training data.
At the positions that enter the supervision, the labels differ from MFA by 53 ms on average, below the 0.1 s resolution of the STM. The label error is thus roughly fifty times smaller than the error of the generated markers (Table R5-2EfV). The gap between marker and audio comes from the model's prediction of its own timing, not from noisy labels.
Q1: Decomposition of the duration error, with numbers
Planning error and realization error. For each reply we measure the realized audio against the instructed duration, the realized audio against the final STM (realization), and the final STM against the instructed duration (planning).
Table R3-2EfV. Decomposition of the duration error (Qwen2.5-Omni-7B + TiCo; TiCo-Bench, 2,000 prompts; MAE in seconds). The first two columns are the two curves of Figure 9, here over all four tasks.
Up to 40 s, planning is nearly exact (the plan is within 1 s of the target in 97–98% of replies), so the error comes almost entirely from realization, as Figure 9 suggests. Beyond 40 s the two curves of Figure 9 still coincide, but the numbers show that planning is no longer exact: the final STM undershoots the target, and the audio running longer than the marker partly offsets it. Stage 2 samples targets only up to 40 s (Appendix C), so this range is outside training.
The final STM alone. Following Table 8, we report the error between the final STM and the ASR timestamp at the same position.
Table R4-2EfV. Alignment error of the final STM (MAE in seconds).
The final marker has a larger absolute error than the all-marker average of Table 8, because timing error accumulates along the reply. In relative terms it is 13.8% of the elapsed time, consistent with the realization error above.
Q2: Cross-validation with an independent aligner
We repeated the analysis of Appendix F.1 with MFA in place of Whisper.
Table R5-2EfV. Marker error against two aligners (Qwen2.5-Omni-7B + TiCo, 320 replies; MAE in seconds).
The marker error is the same against either aligner, and the two aligners differ from each other by 0.08 s. The reported accuracy therefore reflects acoustic timing, not agreement with Whisper. Together with Table R2-2EfV, neither the training labels nor the evaluation depends on the choice of aligner.