Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
FastLongSpeech: Enhancing Large Speech-Language Models for Efficient Long-Speech Processing
Authors: Shoutao Guo, Shaolei Zhang, Qingkai Fang, Zhengrui Ma, Min zhang, Yang Feng
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show that our method exhibits strong performance in both long-speech and short-speech tasks, while greatly improving inference efficiency 2. ... To assess the long-speech capabilities of LSLMs, we develop a long-speech understanding benchmark called Long Speech-Eval. Experiments show that Fast Long Speech achieves efficient speech processing on both long-speech and short-speech benchmarks, and can balance efficiency and effectiveness to meet different requirements. |
| Researcher Affiliation | Academia | 1Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences (ICT/CAS) 2 Key Laboratory of AI Safety, Chinese Academy of Sciences 3 University of Chinese Academy of Sciences, Beijing, China 4 School of Future Science and Engineering, Soochow University EMAIL,EMAIL,EMAIL |
| Pseudocode | No | The paper describes the 'iterative fusion strategy' and 'dynamic compression training method' in detail, along with mathematical formulations (Eq. 1-7) and architectural diagrams (Figure 1), but it does not present a clearly labeled pseudocode or algorithm block. |
| Open Source Code | Yes | 2The Code is at https://github.com/ictnlp/Fast Long Speech.git. |
| Open Datasets | Yes | For the training data in the first stage, we utilize the ASR data, which contain 960 hours of Libri Speech [33] data and 3k hours of data sampled from MLS [34]. In the second training stage, our training data primarily originates from three datasets following the Spoken QA format: Open ASQA [35], Libri SQA [36], and Common Voice [37]. ... For evaluation, we employ a diverse set of nine datasets spanning five distinct tasks... For the ASR task, we use the Libri Speech [33] test-clean and test-other, and Giga Speech [41] test set to evaluate ASR performance. |
| Dataset Splits | Yes | For the training data in the first stage, we utilize the ASR data, which contain 960 hours of Libri Speech [33] data and 3k hours of data sampled from MLS [34]. In the second training stage, our training data primarily originates from three datasets following the Spoken QA format: Open ASQA [35], Libri SQA [36], and Common Voice [37]. ... For evaluation, we employ a diverse set of nine datasets spanning five distinct tasks... For the ASR task, we use the Libri Speech [33] test-clean and test-other, and Giga Speech [41] test set to evaluate ASR performance. ... The Libri SQA test set includes 2620 samples. For the Libri TTS test subset, we select samples corresponding to the Libri TTS test-clean set from Open ASQA, keeping only the 417 samples with a speech duration longer than 15 seconds as our test set. All test sets are under 30s in duration. |
| Hardware Specification | Yes | Table 3 and 4 present the results of inference efficiency experiments, which are obtained on NVIDIA L40. |
| Software Dependencies | No | In the first stage of training for our Fast Long Speech, we only train the CTC decoder. In the second stage, we experimentally set L to {750, 400, 200, 100, 50, 25, 12} and fine-tune the LLM of Qwen2-Audio using Lo RA [45]. ... We use the Sentence Piece10 toolkit to construct the vocabulary for the training of the CTC decoder. ... Both training stages leverage Deep Speed11 Ze RO-2 for optimization. |
| Experiment Setup | Yes | We use Qwen2-Audio-7B-Instruct6 as the base LSLM for all methods, with the length of speech window as 750. Besides, we also extend our method to vanilla Qwen2.5-Omni [15] without dynamic compression training to verify the effectiveness of our method. In the first stage of training for our Fast Long Speech, we only train the CTC decoder. In the second stage, we experimentally set L to {750, 400, 200, 100, 50, 25, 12} and fine-tune the LLM of Qwen2-Audio using Lo RA [45]. For a fair comparison, we also fine-tune Qwen2-Audio for all methods except the Baseline and Fast Long Speech, using the same training data and implementation settings as used for Fast Long Speech. All methods employ the original prompt template from Qwen2-Audio. For more training details, please refer to the Appendix C. ... Table 6: Settings of Fast Long Speech. CTC Decoder: Model hidden_dim 4096, output_dim 10000. Training Details: per_device_batch_size 16, learning_rate 2e-5, lr_scheduler cosine. Base_model: Qwen2-Audio-7B-Instruct, lora_r 128, lora_alpha 256, lora_dropout 0.05, lora_target_modules q_proj, k_proj, v_proj, o_proj. Training Details: per_device_batch_size 16, learning_rate 2e-4, lr_scheduler cosine. |