PDO_S2TT
Official checkpoints for the ICASSP 2027 submission:
PERSISTENT DELIVERY OPTIMIZATION FOR STREAMING SPEECH-TO-TEXT TRANSLATION WITH REVISIONS
PDO is a history-conditioned, direct streaming speech-to-text translation policy. It receives a growing English speech prefix together with the previously displayed translation and may wait, append, or revise the complete visible target-language draft.
Inference, evaluation, and PDO training code
Released checkpoints
| File | Description |
|---|---|
pdo_s2tt.pt |
Inference-only LoRA and history-conditioning weights |
pdo_s2tt_sft.pt |
SFT initialization for reproducing PDO reinforcement learning |
Both checkpoints use Qwen/Qwen3-ASR-1.7B as the base model. The public repository includes the complete multilingual PDO reinforcement-learning stage, its frozen FLEURS training targets, evaluation code, and a four-GPU smoke test. The preceding supervised-training stages and optimizer state are not included.
Supported directions
- En->Zh
- En->De
- En->Es
- En->Ja
- En->Fr
Download
The paper release is pinned to revision 62c27b03aa1d7da1b739e43e815796029ee243f5:
hf download hf-wzx1205/PDO_S2TT pdo_s2tt.pt \
--revision 62c27b03aa1d7da1b739e43e815796029ee243f5 \
--local-dir checkpoints
hf download hf-wzx1205/PDO_S2TT pdo_s2tt_sft.pt \
--revision 62c27b03aa1d7da1b739e43e815796029ee243f5 \
--local-dir checkpoints
See the GitHub training instructions for the released G=4 recipe and complete reproduction workflow.
FLEURS TEST results
| Direction | BLEU | COMET | chrF++ | FTL | FRD | LAAL-CU mean / P90 | RTF |
|---|---|---|---|---|---|---|---|
| En->Zh | 38.16 | 86.31 | 26.92 | 2.00 | 2.33 | 3.53 / 6.61 | .39 |
| En->De | 29.57 | 84.22 | 57.03 | 2.00 | 2.42 | 3.08 / 5.65 | .57 |
| En->Es | 23.35 | 83.57 | 50.88 | 2.00 | 2.44 | 2.51 / 4.44 | .57 |
| En->Ja | 28.66 | 88.50 | 26.47 | 2.00 | 2.47 | 3.67 / 6.23 | .57 |
| En->Fr | 38.16 | 84.26 | 62.06 | 2.00 | 2.44 | 2.46 / 4.93 | .59 |
| Macro | 31.58 | 85.37 | 44.67 | 2.00 | 2.42 | 3.05 / 5.66 | .54 |
FRD and RTF include measured computation; the values above were measured on one RTX 4090. All five directions have 100% final-output coverage.
Intended use
These checkpoints are released for research and reproducibility in revision-capable simultaneous speech translation. They should be used with the accompanying inference and evaluation code. Performance outside the five released English-to-X directions has not been established.
License
The released PDO_S2TT checkpoints are licensed under the Apache License 2.0. The separately distributed Qwen/Qwen3-ASR-1.7B base model is also released under Apache-2.0. FLEURS data and references remain subject to their original dataset licenses.
Model tree for hf-wzx1205/PDO_S2TT
Base model
Qwen/Qwen3-ASR-1.7B