LongYAAL: Latency Metric for Continuous SimultST

• LongYAAL is a long-form, time-based latency metric designed for continuous speech-to-text translation that accounts for inter-segment spillover and eliminates segmentation-induced bias.
• It extends YAAL by applying a global cutoff on the full audio stream, including only tokens emitted before the session’s end to reflect realistic streaming output.
• Experimental evaluations, enhanced by the SoftSegmenter tool, show LongYAAL achieving up to 99% latency ranking accuracy in IWSLT SimulST pipelines.

Searching arXiv for LongYAAL and the cited latency-metrics paper to ground the article in current papers. arXivSearch({"query":"LongYAAL latency simultaneous speech translation", "max_results": 10}) arXivSearch({"query":"(Polák et al., 22 Sep 2025) Better Late Than Never Evaluation of Latency Metrics for Simultaneous Speech-to-Text Translation", "max_results": 5}) LongYAAL is the long-form extension of YAAL (“Yet Another Average Lagging”), introduced for latency evaluation in simultaneous speech-to-text translation (SimulST) on continuous, unsegmented audio streams. It is designed to remove segmentation-induced bias that affects earlier latency metrics and to support realistic streaming evaluation in which system output is produced over an entire session rather than over oracle-presegmented sentences. In current usage, LongYAAL is reported in seconds, often in both computational-unaware (CU) and computational-aware (CA) variants, and has become a central latency measure in recent IWSLT SimulST evaluation pipelines (Polák et al., 22 Sep 2025).

1. Concept and scope

LongYAAL measures the delay between speech input and translated output in long-form SimulST. The metric was introduced after a comprehensive analysis of SimulST latency metrics across language pairs, systems, and both short- and long-form regimes, with the stated goal of addressing a structural bias in current metrics related to segmentation. Its intended setting is continuous audio rather than short, independently processed segments, and its design reflects the fact that realistic streaming systems must operate without oracle sentence boundaries (Polák et al., 22 Sep 2025).

The motivating problem is that many earlier metrics were developed for short-form evaluation and inherit assumptions such as uniform source word duration, no pauses or silence, and monotonic alignment between source and target tokens. In presegmented evaluation, the simulator typically signals “end of source” at the end of each segment and then allows the system to emit the remaining translation tokens immediately. This creates a tail-word artifact: some words appear to incur no delay that would exist in genuine online translation. LongYAAL was introduced to preserve meaningful latency accounting in the presence of such effects by moving the cutoff logic from the segment level to the stream level (Polák et al., 22 Sep 2025).

From the perspective of practical interpretation, LongYAAL behaves as a long-form, time-based latency metric for SimulST that aggregates latency over an entire talk or corpus and yields a single scalar in seconds. This suggests a refined analogue of average lag for speech, adapted to long, unsegmented talks and to contemporary evaluation toolkits used in IWSLT-style setups (Fuxa et al., 2 Jun 2026).

2. Derivation from YAAL and the treatment of tail words

YAAL was proposed as a refinement of LAAL for short-form evaluation. The crucial modification is its cutoff: instead of including tokens up to and including the first one emitted at the end of the segment, YAAL includes only tokens emitted strictly before the end of the source segment. For a segment $\mathbf{X}$ with emission times $d_i$, YAAL defines the cutoff as

$$\tau_{\text{YAAL}}(\mathbf{X}) = \max \{ i \mid d_i < |\mathbf{X}| \}.$$

With the same ideal delay sequence used in LAAL,

$$d_i^* = \frac{i-1}{\gamma}, \qquad \gamma = \frac{\max(|\mathbf{Y}|, |\mathbf{Y}^R|)}{|\mathbf{X}|},$$

YAAL is

$$\text{YAAL}(\mathbf{X}, \mathbf{Y}, \mathbf{Y}^R) = \frac{1}{\tau_{\text{YAAL}}(\mathbf{X})} \sum_{i=1}^{\tau_{\text{YAAL}}(\mathbf{X})} \bigl(d_i - d_i^*\bigr).$$

This definition excludes tail words emitted at or after the end of the segment and thereby makes the metric invariant to whether the simulator requests the remaining translation instantly or not (Polák et al., 22 Sep 2025).

LongYAAL extends that logic to continuous streams. Instead of discarding tokens that spill beyond individual segment boundaries, it includes all hypothesis tokens emitted before the end of the full stream and excludes only the final tail words produced after the end of the entire recording. The paper characterizes this explicitly: LongYAAL includes words generated beyond aligned segment boundaries $\mathbf{X}_s$, but excludes the final tail words produced after the end of the full stream $\mathbf{S}$. If the stream consists of a single segment, LongYAAL coincides with YAAL (Polák et al., 22 Sep 2025).

This shift in cutoff behavior is the core conceptual innovation. In short-form evaluation, segment-level cutoff induces bias because tail behavior is shaped by the segmentation oracle. In long-form evaluation, inter-segment spillover is natural and should remain visible to the metric, whereas only the global offline tail at session end should be removed. That design principle distinguishes LongYAAL from segment-wise adaptations such as StreamLAAL (Polák et al., 22 Sep 2025).

3. Formalization in continuous streams

In the long-form setting, the source is a continuous stream

$$\mathbf{S} = [\mathbf{X}_1, \dots, \mathbf{X}_{|\mathbf{S}|}],$$

with continuous system output

$\mathbf{Y}_S = [y_1, \dots, y_{|\mathbf{Y}_S|}],$

and segmented references $\mathbf{Y}_1^R, \dots, \mathbf{Y}_{|\mathbf{S}|}^R$. Let the total stream duration be

$d_i$0

If $d_i$1 denotes the emission time of hypothesis token $d_i$2, LongYAAL uses the global cutoff

$d_i$3

The stream-level length ratio is

$d_i$4

and the ideal delays are

$d_i$5

LongYAAL is then written as

$d_i$6

This formulation is conceptually equivalent to YAAL but moves the computation from the sentence level to the level of the entire stream. In implementation, references may remain segmented and hypotheses may be resegmented for alignment, but the inclusion and exclusion rule is global: tokens with $d_i$7 are included, and tokens with $d_i$8 are excluded (Polák et al., 22 Sep 2025).

In systems that use word-level timing from ASR alignment, LongYAAL can also be understood operationally through source-word end times and target emission times. In that perspective, latency resembles a word-level lagging term

$d_i$9

where $$\tau_{\text{YAAL}}(\mathbf{X}) = \max \{ i \mid d_i < |\mathbf{X}| \}.$$0 is the end time of the $$\tau_{\text{YAAL}}(\mathbf{X}) = \max \{ i \mid d_i < |\mathbf{X}| \}.$$1-th source word, $$\tau_{\text{YAAL}}(\mathbf{X}) = \max \{ i \mid d_i < |\mathbf{X}| \}.$$2 is the emission time of target word $$\tau_{\text{YAAL}}(\mathbf{X}) = \max \{ i \mid d_i < |\mathbf{X}| \}.$$3, and $$\tau_{\text{YAAL}}(\mathbf{X}) = \max \{ i \mid d_i < |\mathbf{X}| \}.$$4 maps target units to source units. The exact LongYAAL formula remains that of Polák et al.; the latter expression is a generic interpretation of the metric’s underlying logic in systems that expose online word timestamps (Fuxa et al., 2 Jun 2026).

4. Segmentation, alignment, and SoftSegmenter

LongYAAL depends on reliable resegmentation of continuous output against segmented references. For that purpose, the latency-evaluation paper introduces SoftSegmenter, a resegmentation tool intended to support robust long-form evaluation. Its inputs are reference translations, system hypotheses, and hypothesis emission times. Its goal is to assign each hypothesis token to a reference segment, or leave it unassigned at the stream end, while respecting chronology, punctuation type, and token similarity (Polák et al., 22 Sep 2025).

SoftSegmenter lowercases and tokenizes both reference and hypothesis texts; for Chinese and Japanese it operates at character level. For each reference token $$\tau_{\text{YAAL}}(\mathbf{X}) = \max \{ i \mid d_i < |\mathbf{X}| \}.$$5 and hypothesis token $$\tau_{\text{YAAL}}(\mathbf{X}) = \max \{ i \mid d_i < |\mathbf{X}| \}.$$6, it defines an alignment score

$$\tau_{\text{YAAL}}(\mathbf{X}) = \max \{ i \mid d_i < |\mathbf{X}| \}.$$7

where $$\tau_{\text{YAAL}}(\mathbf{X}) = \max \{ i \mid d_i < |\mathbf{X}| \}.$$8 is the start time of the reference segment containing $$\tau_{\text{YAAL}}(\mathbf{X}) = \max \{ i \mid d_i < |\mathbf{X}| \}.$$9, $$d_i^* = \frac{i-1}{\gamma}, \qquad \gamma = \frac{\max(|\mathbf{Y}|, |\mathbf{Y}^R|)}{|\mathbf{X}|},$$0 is the emission time of the hypothesis token, $$d_i^* = \frac{i-1}{\gamma}, \qquad \gamma = \frac{\max(|\mathbf{Y}|, |\mathbf{Y}^R|)}{|\mathbf{X}|},$$1 indicates punctuation, and the similarity term is

$$d_i^* = \frac{i-1}{\gamma}, \qquad \gamma = \frac{\max(|\mathbf{Y}|, |\mathbf{Y}^R|)}{|\mathbf{X}|},$$2

The chronology constraint prevents aligning a token to a future segment, which is important because such errors can otherwise yield spurious negative latency or unstable assignments (Polák et al., 22 Sep 2025).

Experimentally, the role of SoftSegmenter is substantial. On reconcatenated short-form outputs, replacing mWERSegmenter with SoftSegmenter raises latency-ranking accuracy from approximately 86% to approximately 94%, while MT-quality accuracy remains at least 99% for both. In long-form experiments, direct application of short-form metrics to streams yields only 57–66% latency accuracy, and AP is substantially worse at approximately 39%; with SoftSegmenter-based resegmentation, LongLAAL and LongYAAL reach 95–99% accuracy depending on the subset, while StreamLAAL using mWERSegmenter trails by 8–10 points (Polák et al., 22 Sep 2025).

A practical implication is that LongYAAL is not only a metric definition but part of an evaluation stack. Its reported reliability depends on time-aware resegmentation, and SoftSegmenter provides the mechanism that makes stream-level inclusion rules operational on segmented references.

5. CU and CA LongYAAL in IWSLT 2026 SimulST

In the IWSLT 2026 Simultaneous Speech Translation task, translation latency is reported with LongYAAL in both computational-unaware and computational-aware variants. The evaluation setup described for AlignAtt4LLM uses OmniSTEval and Simulstream, and reports BLEU, chrF, XCOMET-XL, and LongYAAL for English to German, Italian, and Chinese (Fuxa et al., 2 Jun 2026).

The two LongYAAL variants differ only in the timestamps used during evaluation.

Variant	Timestamp basis	Interpretation
CU-LongYAAL	Fixed chunk cadence	Idealized real-time engine
CA-LongYAAL	Actual wall-clock processing time	Hardware-aware latency

In CU mode, each chunk boundary contributes a fixed increment equal to the chunk size, such as $$d_i^* = \frac{i-1}{\gamma}, \qquad \gamma = \frac{\max(|\mathbf{Y}|, |\mathbf{Y}^R|)}{|\mathbf{X}|},$$3 ms or $$d_i^* = \frac{i-1}{\gamma}, \qquad \gamma = \frac{\max(|\mathbf{Y}|, |\mathbf{Y}^R|)}{|\mathbf{X}|},$$4 ms. In CA mode, each chunk boundary contributes the actual wall-clock time spent processing that chunk. Because the deployed system runs faster than real time on a single NVIDIA A40, CA-LongYAAL is consistently below CU-LongYAAL in the reported experiments (Fuxa et al., 2 Jun 2026).

The submission defines two official regimes directly characterized by CU-LongYAAL.

Regime	Chunk size	CU-LongYAAL target
Low-latency	$$d_i^* = \frac{i-1}{\gamma}, \qquad \gamma = \frac{\max(|\mathbf{Y}|, |\mathbf{Y}^R|)}{|\mathbf{X}|},$$5 ms	Around $$d_i^* = \frac{i-1}{\gamma}, \qquad \gamma = \frac{\max(|\mathbf{Y}|, |\mathbf{Y}^R|)}{|\mathbf{X}|},$$6 s
High-latency	$$d_i^* = \frac{i-1}{\gamma}, \qquad \gamma = \frac{\max(|\mathbf{Y}|, |\mathbf{Y}^R|)}{|\mathbf{X}|},$$7 ms	Below $$d_i^* = \frac{i-1}{\gamma}, \qquad \gamma = \frac{\max(|\mathbf{Y}|, |\mathbf{Y}^R|)}{|\mathbf{X}|},$$8 s

Both regimes also delay the first MT emission until $$d_i^* = \frac{i-1}{\gamma}, \qquad \gamma = \frac{\max(|\mathbf{Y}|, |\mathbf{Y}^R|)}{|\mathbf{X}|},$$9 seconds of source audio have accumulated. The system architecture is a synchronous cascade in which Qwen3-ASR with forced alignment produces an incrementally updated source transcript, and Gemma-4 E4B-it translates that prefix under an MT-side AlignAtt policy. In each chunk, ASR retranscribes the live tail, the forced aligner assigns word times, the transcript prefix is stabilized up to punctuation, and one MT request is launched on the updated transcript. This strict alternation simplifies CA measurement because there is no overlapping ASR and MT execution (Fuxa et al., 2 Jun 2026).

LongYAAL is the primary latency objective against which the cascade is tuned. The AlignAtt policy imposes a source-frontier constraint,

$$\text{YAAL}(\mathbf{X}, \mathbf{Y}, \mathbf{Y}^R) = \frac{1}{\tau_{\text{YAAL}}(\mathbf{X})} \sum_{i=1}^{\tau_{\text{YAAL}}(\mathbf{X})} \bigl(d_i - d_i^*\bigr).$$0

with $$\text{YAAL}(\mathbf{X}, \mathbf{Y}, \mathbf{Y}^R) = \frac{1}{\tau_{\text{YAAL}}(\mathbf{X})} \sum_{i=1}^{\tau_{\text{YAAL}}(\mathbf{X})} \bigl(d_i - d_i^*\bigr).$$1, so a draft token is not committed if its reconstructed alignment lies beyond the accessible source frontier. Additional gates based on argmax mass and source-mass thresholds are present but inactive in the official runs, with $$\text{YAAL}(\mathbf{X}, \mathbf{Y}, \mathbf{Y}^R) = \frac{1}{\tau_{\text{YAAL}}(\mathbf{X})} \sum_{i=1}^{\tau_{\text{YAAL}}(\mathbf{X})} \bigl(d_i - d_i^*\bigr).$$2 and $$\text{YAAL}(\mathbf{X}, \mathbf{Y}, \mathbf{Y}^R) = \frac{1}{\tau_{\text{YAAL}}(\mathbf{X})} \sum_{i=1}^{\tau_{\text{YAAL}}(\mathbf{X})} \bigl(d_i - d_i^*\bigr).$$3. Offline head selection retains the top $$\text{YAAL}(\mathbf{X}, \mathbf{Y}, \mathbf{Y}^R) = \frac{1}{\tau_{\text{YAAL}}(\mathbf{X})} \sum_{i=1}^{\tau_{\text{YAAL}}(\mathbf{X})} \bigl(d_i - d_i^*\bigr).$$4 attention heads per language pair by Translation Score, reducing reconstruction cost and improving latency. In the head-filtering ablation, Top-8 heads yield XCOMET-XL $$\text{YAAL}(\mathbf{X}, \mathbf{Y}, \mathbf{Y}^R) = \frac{1}{\tau_{\text{YAAL}}(\mathbf{X})} \sum_{i=1}^{\tau_{\text{YAAL}}(\mathbf{X})} \bigl(d_i - d_i^*\bigr).$$5, CU $$\text{YAAL}(\mathbf{X}, \mathbf{Y}, \mathbf{Y}^R) = \frac{1}{\tau_{\text{YAAL}}(\mathbf{X})} \sum_{i=1}^{\tau_{\text{YAAL}}(\mathbf{X})} \bigl(d_i - d_i^*\bigr).$$6 s, and CA $$\text{YAAL}(\mathbf{X}, \mathbf{Y}, \mathbf{Y}^R) = \frac{1}{\tau_{\text{YAAL}}(\mathbf{X})} \sum_{i=1}^{\tau_{\text{YAAL}}(\mathbf{X})} \bigl(d_i - d_i^*\bigr).$$7 s, whereas using all $$\text{YAAL}(\mathbf{X}, \mathbf{Y}, \mathbf{Y}^R) = \frac{1}{\tau_{\text{YAAL}}(\mathbf{X})} \sum_{i=1}^{\tau_{\text{YAAL}}(\mathbf{X})} \bigl(d_i - d_i^*\bigr).$$8 heads yields XCOMET-XL $$\text{YAAL}(\mathbf{X}, \mathbf{Y}, \mathbf{Y}^R) = \frac{1}{\tau_{\text{YAAL}}(\mathbf{X})} \sum_{i=1}^{\tau_{\text{YAAL}}(\mathbf{X})} \bigl(d_i - d_i^*\bigr).$$9, CU $\mathbf{X}_s$0 s, and CA $\mathbf{X}_s$1 s (Fuxa et al., 2 Jun 2026).

The end-to-end development-set results show how LongYAAL supports quality–latency trade-off analysis. For English to German, the low-latency baseline has BLEU $\mathbf{X}_s$2, XCOMET $\mathbf{X}_s$3, and CU-LongYAAL $\mathbf{X}_s$4 s, whereas AlignAtt4LLM reports BLEU $\mathbf{X}_s$5, XCOMET $\mathbf{X}_s$6, CU-LongYAAL $\mathbf{X}_s$7 s, and CA-LongYAAL $\mathbf{X}_s$8 s. In the high-latency regime, the baseline has BLEU $\mathbf{X}_s$9, XCOMET $\mathbf{S}$0, and CU-LongYAAL $\mathbf{S}$1 s, whereas AlignAtt4LLM reports BLEU $\mathbf{S}$2, XCOMET $\mathbf{S}$3, CU-LongYAAL $\mathbf{S}$4 s, and CA-LongYAAL $\mathbf{S}$5 s. For English to Italian, the system likewise exceeds the baseline in both regimes, with low-latency values of BLEU $\mathbf{S}$6, XCOMET $\mathbf{S}$7, CU $\mathbf{S}$8 s, and CA $\mathbf{S}$9 s, and high-latency values of BLEU $$\mathbf{S} = [\mathbf{X}_1, \dots, \mathbf{X}_{|\mathbf{S}|}],$$0, XCOMET $$\mathbf{S} = [\mathbf{X}_1, \dots, \mathbf{X}_{|\mathbf{S}|}],$$1, CU $$\mathbf{S} = [\mathbf{X}_1, \dots, \mathbf{X}_{|\mathbf{S}|}],$$2 s, and CA $$\mathbf{S} = [\mathbf{X}_1, \dots, \mathbf{X}_{|\mathbf{S}|}],$$3 s. For English to Chinese, the results are more mixed: in the low-latency regime, the baseline has BLEU $$\mathbf{S} = [\mathbf{X}_1, \dots, \mathbf{X}_{|\mathbf{S}|}],$$4, XCOMET $$\mathbf{S} = [\mathbf{X}_1, \dots, \mathbf{X}_{|\mathbf{S}|}],$$5, and CU $$\mathbf{S} = [\mathbf{X}_1, \dots, \mathbf{X}_{|\mathbf{S}|}],$$6 s, whereas the AlignAtt4LLM system has BLEU $$\mathbf{S} = [\mathbf{X}_1, \dots, \mathbf{X}_{|\mathbf{S}|}],$$7, XCOMET $$\mathbf{S} = [\mathbf{X}_1, \dots, \mathbf{X}_{|\mathbf{S}|}],$$8, CU $$\mathbf{S} = [\mathbf{X}_1, \dots, \mathbf{X}_{|\mathbf{S}|}],$$9 s, and CA $\mathbf{Y}_S = [y_1, \dots, y_{|\mathbf{Y}_S|}],$0 s; in the high-latency regime, the baseline has BLEU $\mathbf{Y}_S = [y_1, \dots, y_{|\mathbf{Y}_S|}],$1, XCOMET $\mathbf{Y}_S = [y_1, \dots, y_{|\mathbf{Y}_S|}],$2, and CU $\mathbf{Y}_S = [y_1, \dots, y_{|\mathbf{Y}_S|}],$3 s, whereas the system has BLEU $\mathbf{Y}_S = [y_1, \dots, y_{|\mathbf{Y}_S|}],$4, XCOMET $\mathbf{Y}_S = [y_1, \dots, y_{|\mathbf{Y}_S|}],$5, CU $\mathbf{Y}_S = [y_1, \dots, y_{|\mathbf{Y}_S|}],$6 s, and CA $\mathbf{Y}_S = [y_1, \dots, y_{|\mathbf{Y}_S|}],$7 s (Fuxa et al., 2 Jun 2026).

6. Relation to earlier metrics, interpretation, and limitations

LongYAAL was proposed in direct comparison with AP, AL, LAAL, DAL, ATD, StreamLAAL, and related long-form adaptations. The evaluation paper concludes that YAAL and LongYAAL outperform popular latency metrics, while SoftSegmenter enhances alignment quality in long-form evaluation. In short-form, YAAL attains approximately 96% ranking accuracy over all system pairs and up to approximately 99% after filtering nearly equal-latency pairs. In long-form with SoftSegmenter-based resegmentation, LongYAAL is reported at approximately 95% accuracy over all pairs and approximately 99% accuracy on pairs with clearly different latencies, placing it at the top or tied for top among the tested long-form metrics (Polák et al., 22 Sep 2025).

The main distinction from StreamLAAL is methodological. StreamLAAL computes LAAL per segment after resegmentation and averages over segments, which preserves a segment-level cutoff and inherits tail-related bias. LongYAAL instead uses a stream-level cutoff, allows tokens to extend beyond aligned segment boundaries, and excludes only the global tail after the end of the full recording. The paper characterizes LongLAAL as slightly less robust in regimes with large latency differences, likely because it discards tokens beyond reference segment boundaries (Polák et al., 22 Sep 2025).

Several limitations remain explicit in the literature. LongYAAL still relies on a global $\mathbf{Y}_S = [y_1, \dots, y_{|\mathbf{Y}_S|}],$8 based on length ratios, which assumes roughly uniform speech speed and translation density across the stream. LongYAAL itself requires emission times and reference tokens, and while true-latency validation additionally uses forced alignment and bilingual word alignment, the metric remains dependent on the quality of resegmentation. SoftSegmenter improves alignment reliability, but strong ASR errors, disfluencies, or domain mismatch can still introduce noise. The authors also note that latency metrics can be gamed and recommend combining them with qualitative and human evaluation (Polák et al., 22 Sep 2025).

In system-design practice, LongYAAL functions as an anchor metric for quality–latency optimization. In the IWSLT 2026 cascade described above, chunk size, synchronization policy, alignment-head selection, and frontier gating are all tuned to operate at CU-LongYAAL targets around $\mathbf{Y}_S = [y_1, \dots, y_{|\mathbf{Y}_S|}],$9 seconds and below $\mathbf{Y}_1^R, \dots, \mathbf{Y}_{|\mathbf{S}|}^R$0 seconds. This suggests that LongYAAL is not merely descriptive: it shapes architectural choices, scheduler choices, and the calibration of online commitment policies in contemporary SimulST systems (Fuxa et al., 2 Jun 2026).