In streaming simultaneous speech translation, the speech translation system is trained to learn a read-write policy that alternates between consuming source words and generating target ones. In a cascaded setting, the output from the speech recognizer is passed to a separate machine translation component, making it more difficult to learn such a policy. Approximations such as fixed wait-k strategies or target-suffix deletion can be employed, but these approaches do not provide the model with a streaming system's flexibility to make contextual read-write decisions. This paper presents a training strategy for a cascaded machine translation system that enables it to dynamically decide how much of the growing source prefix to translate. We achieve this by fine-tuning a large language model (Qwen3-8B) on stable prefixes of the training data, which are produced by pairing every source sentence prefix in the training data with the longest translation of that prefix that is shared with the full source sentence translation. We fine-tune variants of the model on different subsets of the prefixes and compare against wait-k and target-suffix deletion. We also investigate the effect of fine-tuning the target-token generation confidence. Our experiments show that stable prefixes improve the quality-latency tradeoff when translating from English into German, Japanese, and Chinese across a range of test sets.
Figures & tables
#
type
source prefix
translation
stable prefix
final
Does he playfully tease you?
Neckt er dich spielerisch?
Neckt er dich spielerisch?
1
partial
Does
Macht
2
partial
Does he
Tut er das
3
partial
Does he playfully
Tut er das spielerisch
4
partial
Does he playfully tease
Neckt er
Neckt er
Table 1: Stable-prefix extraction for a German example. Not until step 4 is a partial translation a prefix of the full-sentence translation.
ID
model
AL
Full
Worst
EN → DE
1
base
0.82
85.9
80.9
2
prefix
2.21
87.1
85.1
3
continuation
2.21
86.1
81.2
EN → JA
4
base
0.11
89.2
81.3
Table 2: Average Lagging and COMET ( ×100 ) on FLEURS. Full is complete-sentence COMET; Worst is the minimum corpus-average COMET over source boundaries.
Figure 1: Binary source-completion on FLEURS: COMET of the recomposed translation (top) and prefix-translation length (bottom) versus source words revealed, for the pre-trained base and the prefix adapter prefix .
wait- k
suffix del.
confidence
base
EN → DE
47.9
75.1
75.0
EN → JA
76.8
82.7
85.6
EN → ZH
81.1
84.5
86.2
prefix
EN → DE
65.0
73.7
78.9
Table 3: Latency-mechanism AUC on FLEURS summarizing Figure 2 : COMET over Average Lagging. Bold : best mechanism per row; ⋆ : best per language.
Figure 2: The three latency mechanisms per model (rows) × EN → {DE,JA,ZH} (columns) on FLEURS.
Figure 3: Expected Calibration Error (lower is better) of normalized commit confidence qgo against the FLEURS stable-prefix oracle. Bars are 95% intervals from 1,000 sentence-level bootstrap samples. Top: all non-final source prefixes; bottom: prefixes with at most 30% of the source visible.
Figure 4: Confidence quality–latency frontiers on FLEURS for base , prefix , continuation , and the full-control adapter.
base
ft. + conf.
wait- k
suf. del.
conf.
prefix
cont.
FLEURS
DE
48.2
75.1
75.1
79.0
81.9
JA
77.0
82.7
85.6
87.3
86.5
ZH
81.1
84.5
86.2
87.0
86.5
WMT24++
Table 4: COMET quality–latency frontier AUC summarizing Figure 5 over Average Lagging, for EN → {DE,JA,ZH}. Each column scores one sweep on the shared latency interval for that dataset–language panel; best per row in bold .
Figure 5: COMET quality–latency tradeoff on FLEURS, WMT24++, and CoVoST 2 (rows) × EN → {DE,JA,ZH} (columns). Each panel compares base under wait- k , suffix deletion, and confidence with prefix and continuation under confidence.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Direction
prefix
continuation
EN → DE
2,000 / 4,500 (576k)
3,500 / 6,000 (768k)
EN → JA
1,000 / 3,500 (448k)
10,500 / 13,000 (1.664M)
EN → ZH
2,000 / 4,500 (576k)
8,500 / 11,000 (1.408M)
Appendix
Table 5: Selected update / final update evaluated before stopping; parentheses give the total number of optimized records through the final update (updates multiplied by effective batch size 128).
base
ft. + conf.
wait- k
suf. del.
conf.
prefix
cont.
FLEURS
DE
10.4
4.2
4.5
3.1
2.5
JA
7.7
6.2
5.1
4.3
4.6
ZH
4.0
3.0
2.6
2.2
2.3
WMT24++
Appendix
Table 6: MetricX-error quality–latency frontier AUC corresponding to Table 4 (lower is better), for EN → {DE,JA,ZH}. Each column scores one sweep on the shared latency interval for that dataset–language panel; best per row in bold .
Figure 6: MetricX version of Figure 5 , with the same five sweeps (MetricX error, lower is better ).