Unified Segment-to-Segment Framework for Simultaneous Sequence Generation

Authors: Shaolei Zhang, Yang Feng

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on multiple simultaneous generation tasks demonstrate that Seg2Seg achieves state-of-the-art performance and exhibits better generality across various tasks2.
Researcher Affiliation Academia 1Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences (ICT/CAS) 2University of Chinese Academy of Sciences
Pseudocode Yes Algorithm 1 illustrates the specific inference process of Seg2Seg.
Open Source Code Yes 2Code is available at: https://github.com/ictnlp/Seg2Seg.
Open Datasets Yes We apply Libri Speech3 benchmark [59], which consists of 960 hours English audio.
Dataset Splits Yes We use dev-clean (5.4 hours) and dev-other (5.3 hours) as validation sets, and test-clean (5.4 hours) and test-other (5.1 hours) as test sets, where test-other set contains more noisy audio. For speech, we use the raw 16-bit 16k Hz mono-channel audio wave. For text, we use Sentence Piece [60] to generate a unigram vocabulary of size 10000.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU models, CPU types, memory) used for running the experiments. It only mentions using a Transformer-Base model and pre-trained Wav2Vec2.0.
Software Dependencies No The paper mentions software like Fairseq Library [66], Wav2Vec2.0 [67], and Simul Eval [68], but it does not specify exact version numbers for these software dependencies, which is required for reproducibility.
Experiment Setup Yes In Seg2Seg, we use the standard Transformer-Base (6 encoder and 6 decoder layers) [55] for Simul MT. For streaming ASR and Simul ST, we replace the word embedding layer in Transformer-Base with a pre-trained Wav2Vec2.06 [67] to extract the acoustic embedding, and the rest remains the same as Simul MT.