lcma-srt

LCMA-SRT: Language-Conditional Mixture-of-Experts Adapters for Joint Multilingual Speech Recognition and Translation

Neural transducers offer an alignment-free framework for speech-to-text modeling, and hierarchical transducer architectures further improve multilingual joint automatic speech recognition (ASR) and speech translation (ST) by stacking a translation-focused encoder on top of an ASR encoder. However, extending hierarchical transducers to multilingual many-to-many settings remains challenging: fully shared models often suffer from negative transfer and unstable target-language generation, while training separate models for each direction is computationally prohibitive. We propose LCMA-SRT (Language-Conditional Mixture-of-Experts Adapters for Speech Recognition and Translation), which augments a hierarchical transducer with language-conditional Mixture-of-Experts (MoE) adapters. A source-conditioned MoE adapter (SRC-MoE) uses source-language embeddings to reduce cross-language interference and improve multilingual ASR. A target-conditioned MoE adapter (TGT-MoE) uses the desired target language to reduce cross-target interference and stabilize target-language generation in many-to-many ST. Experiments on Europarl-ST (9 languages, 72 directions) show that LCMA-SRT improves both ASR and ST within a single joint model, reducing average WER and improving BLEU and COMET over strong hierarchical transducer baselines.

Main Results

Multilingual ASR Pretraining

<table> <thead> <tr> <th rowspan="2">Model</th> <th colspan="10">WER (%) ↓</th> </tr> <tr> <th>de</th><th>en</th><th>es</th><th>fr</th><th>it</th> <th>nl</th><th>pl</th><th>pt</th><th>ro</th><th>Avg</th> </tr> </thead> <tbody> <tr> <td>CR-CTC</td> <td>24.57</td><td>18.59</td><td>20.76</td><td>19.24</td><td>17.33</td> <td>36.75</td><td>25.28</td><td>19.82</td><td>18.77</td><td>22.35</td> </tr> <tr> <td>&nbsp;&nbsp;&nbsp;+ MoE</td> <td>24.39</td><td>18.41</td><td>20.16</td><td>18.61</td><td>17.28</td> <td>36.83</td><td>24.36</td><td>19.70</td><td>18.79</td><td>22.06</td> </tr> <tr> <td>&nbsp;&nbsp;&nbsp;+ S-Bias</td> <td>23.89</td><td>17.60</td><td>19.58</td><td>17.41</td><td>16.73</td> <td><strong>34.72</strong></td><td>23.63</td><td>18.21</td><td>17.97</td><td>21.08</td> </tr> <tr> <td>&nbsp;&nbsp;&nbsp;+ SRC-MoE</td> <td><strong>23.34</strong></td><td><strong>17.45</strong></td><td><strong>19.41</strong></td> <td><strong>17.34</strong></td><td><strong>16.27</strong></td><td>35.20</td> <td><strong>23.28</strong></td><td><strong>18.16</strong></td><td><strong>17.48</strong></td> <td><strong>20.88</strong></td> </tr> </tbody> </table>

Many-to-Many Joint Training (Average)

<table> <thead> <tr> <th rowspan="2">Model</th> <th rowspan="2">WER (%)↓</th> <th colspan="10">Average BLEU ↑</th> </tr> <tr> <th>de</th><th>en</th><th>es</th><th>fr</th><th>it</th><th>nl</th><th>pl</th><th>pt</th><th>ro</th><th>Avg</th> </tr> </thead> <tbody> <tr> <td>HENT-SRT-M20×9</td> <td>23.28</td> <td>10.7</td><td>21.2</td><td>19.1</td><td>18.2</td><td>14.2</td><td>16.5</td><td>7.2</td><td>18.4</td><td>12.1</td><td>15.3</td> </tr> <tr> <td>HENT-SRT-M2M</td> <td>16.65</td> <td>2.6</td><td>12.8</td><td>5.5</td><td>4.0</td><td>1.8</td><td>3.5</td><td>1.2</td><td>4.9</td><td>2.5</td><td>4.3</td> </tr> <tr> <td>LCMA-SRT</td> <td><strong>15.71</strong></td> <td><strong>15.2</strong></td><td><strong>25.9</strong></td><td><strong>25.8</strong></td><td><strong>24.7</strong></td> <td><strong>20.0</strong></td><td><strong>20.5</strong></td><td><strong>10.7</strong></td><td><strong>23.9</strong></td><td><strong>17.6</strong></td><td><strong>20.5</strong></td> </tr> </tbody> </table> <table> <thead> <tr> <th rowspan="2">Model</th> <th rowspan="2">LMR (%)↓</th> <th colspan="10">Average COMET ↑</th> </tr> <tr> <th>de</th><th>en</th><th>es</th><th>fr</th><th>it</th><th>nl</th><th>pl</th><th>pt</th><th>ro</th><th>Avg</th> </tr> </thead> <tbody> <tr> <td>HENT-SRT-M20×9</td> <td>0.65</td> <td>0.507</td><td>0.656</td><td>0.587</td><td>0.542</td><td>0.565</td><td>0.558</td><td>0.550</td><td>0.609</td><td>0.598</td><td>0.575</td> </tr> <tr> <td>HENT-SRT-M2M</td> <td>84.95</td> <td>0.380</td><td>0.543</td><td>0.478</td><td>0.427</td><td>0.435</td><td>0.401</td><td>0.385</td><td>0.471</td><td>0.406</td><td>0.436</td> </tr> <tr> <td>LCMA-SRT</td> <td>0.75</td> <td><strong>0.574</strong></td><td><strong>0.715</strong></td><td><strong>0.682</strong></td><td><strong>0.627</strong></td> <td><strong>0.656</strong></td><td><strong>0.613</strong></td><td><strong>0.616</strong></td><td><strong>0.693</strong></td><td><strong>0.678</strong></td><td><strong>0.651</strong></td> </tr> </tbody> </table>

Usage

1. Install dependencies

# Clone the code repository
git clone https://github.com/linanjie0820/LCMA-SRT.git
cd LCMA-SRT

# Install icefall environment
# See: https://k2-fsa.github.io/icefall/installation/index.html
pip install safetensors sentencepiece torch torchaudio kaldifeat k2 lhotse

2. Download model

pip install huggingface_hub
huggingface-cli download linanjie0820/lcma-srt --local-dir ./model

3. Inference

# ASR (speech recognition)
python model/inference.py --audio test.wav --task asr --src-lang en

# AST (speech translation, e.g., English -> German)
python model/inference.py --audio test.wav --task ast --src-lang en --tgt-lang de

# Both ASR + AST simultaneously
python model/inference.py --audio test.wav --task both --src-lang en --tgt-lang de

Example Output

ASR only:

$ python inference.py --audio en_5525.flac --task asr --src-lang en

==================================================
Decoding ASR...
[ASR] mister president one of the key issues in the new lisbon treaty is the increased role of the european union in the world
==================================================

AST only (English -> German):

$ python inference.py --audio en_5525.flac --task ast --src-lang en --tgt-lang de

==================================================
Decoding AST (en -> de)...
[AST en->de] herr präsident eine wichtige des neuen vertrag von lissabon ist die erhöhten der europäischen union in der welt
==================================================

Both ASR + AST:

$ python inference.py --audio en_5525.flac --task both --src-lang en --tgt-lang de

==================================================
Decoding ASR...
[ASR] mister president one of the key issues in the new lisbon treaty is the increased role of the european union in the world
Decoding AST (en -> de)...
[AST en->de] herr präsident eine wichtige des neuen vertrag von lissabon ist die erhöhten der europäischen union in der welt
==================================================

Loading weights in Python

from safetensors.torch import load_file

# Load weights
state_dict = load_file("model/lcma-srt.safetensors")

# Build model (requires LCMA-SRT code repo)
model.load_state_dict(state_dict, strict=False)
model.eval()

Files

Training & Decoding

Supported Languages

The dataset covers 9 European languages:

All non-identical language pairs are processed (72 pairs total).

Output Format

Each final CutSet manifest entry (MonoCut) contains:

{
  "id": "en_5525-758",
  "start": 0,
  "duration": 6.76,
  "channel": 0,
  "supervisions": [
    {
      "id": "en_5525",
      "recording_id": "en_5525",
      "start": 0.0,
      "duration": 6.76,
      "channel": 0,
      "text": "mister president one of the key issues in the new lisbon treaty is the increased role of the european union in the world",
      "language": "en",
      "speaker": "unknown",
      "custom": {
        "st_text": "herr präsident einer der hauptpunkte des neuen vertrages von lissabon ist die aufwertung der rolle der europäischen union in der welt",
        "lang": "de"
      }
    }
  ],
  "recording": {
    "id": "en_5525",
    "sources": [
      {
        "type": "file",
        "channels": [0],
        "source": "data/Europarl-ST/audio/test/en_5525.flac"
      }
    ],
    "sampling_rate": 16000,
    "num_samples": 108160,
    "duration": 6.76
  },
  "type": "MonoCut"
}

Key fields:

• supervisions[0].text: ASR ground truth transcription
• supervisions[0].language: source language
• supervisions[0].custom.st_text: ST ground truth translation
• supervisions[0].custom.lang: target language

For training and full decoding scripts, please refer to the GitHub repository.

Citation

@inproceedings{li2026lcma,
  title={LCMA-SRT: Language-Conditional Mixture-of-Experts Adapters for Joint Multilingual Speech Recognition and Translation},
  author={Li, Nanjie and Guo, Xiaoyong and Huang, Hao and Haihua, Xu and Shi, Wei},
  booktitle={Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  pages={35363--35377},
  year={2026}
}

License

Apache License 2.0