The paper reports on a series of experiments aiming at probing LeBenchmark, a pretrained acoustic model trained on 7k hours of spoken French, for syntactic information. Pretrained acoustic models are increasingly used for downstream speech tasks such as automatic speech recognition, speech translation, spoken language understanding or speech parsing. They are trained on very low level information (the raw speech signal), and do not have explicit lexical knowledge. Despite that, they obtained reasonable results on tasks that requires higher level linguistic knowledge. As a result, an emerging question is whether these models encode syntactic information. We probe each representation layer of LeBenchmark for syntax, using the Orf\'eo treebank, and observe that it has learnt some syntactic information. Our results show that syntactic information is more easily extractable from the middle layers of the network, after which a very sharp decrease is observed.
Audio-visual speech recognition (AVSR) is a multimodal extension of automatic speech recognition (ASR), using video as a complement to audio. In AVSR, considerable efforts have been directed at datasets for facial features such as lip-readings, while they often fall short in evaluating the image comprehension capabilities in broader contexts. In this paper, we construct SlideAVSR, an AVSR dataset using scientific paper explanation videos. SlideAVSR provides a new benchmark where models transcribe speech utterances with texts on the slides on the presentation recordings. As technical terminologies that are frequent in paper explanations are notoriously challenging to transcribe without reference texts, our SlideAVSR dataset spotlights a new aspect of AVSR problems. As a simple yet effective baseline, we propose DocWhisper, an AVSR model that can refer to textual information from slides, and confirm its effectiveness on SlideAVSR.
It has been shown that the intelligibility of noisy speech can be improved by speech enhancement (SE) algorithms. However, monaural SE has not been established as an effective frontend for automatic speech recognition (ASR) in noisy conditions compared to an ASR model trained on noisy speech directly. The divide between SE and ASR impedes the progress of robust ASR systems, especially as SE has made major advances in recent years. This paper focuses on eliminating this divide with an ARN (attentive recurrent network) time-domain and a CrossNet time-frequency domain enhancement models. The proposed systems fully decouple frontend enhancement and backend ASR trained only on clean speech. Results on the WSJ, CHiME-2, LibriSpeech, and CHiME-4 corpora demonstrate that ARN and CrossNet enhanced speech both translate to improved ASR results in noisy and reverberant environments, and generalize well to real acoustic scenarios. The proposed system outperforms the baselines trained on corrupted speech directly. Furthermore, it cuts the previous best word error rate (WER) on CHiME-2 by $28.4\%$ relatively with a $5.57\%$ WER, and achieves $3.32/4.44\%$ WER on single-channel CHiME-4 simulated/real test data without training on CHiME-4.
There is a growing interest in cost-effective self-supervised fine-tuning (SSFT) of self-supervised learning (SSL)-based speech models to obtain task-specific representations. These task-specific representations are used for robust performance on various downstream tasks by fine-tuning on the labelled data. This work presents a cost-effective SSFT method named Self-supervised Correspondence (SCORE) fine-tuning to adapt the SSL speech representations for content-related tasks. The proposed method uses a correspondence training strategy, aiming to learn similar representations from perturbed speech and original speech. Commonly used data augmentation techniques for content-related tasks (ASR) are applied to obtain perturbed speech. SCORE fine-tuned HuBERT outperforms the vanilla HuBERT on SUPERB benchmark with only a few hours of fine-tuning (< 5 hrs) on a single GPU for automatic speech recognition, phoneme recognition, and query-by-example tasks, with relative improvements of 1.09%, 3.58%, and 12.65%, respectively. SCORE provides competitive results with the recently proposed SSFT method SPIN, using only 1/3 of the processed speech compared to SPIN.
Endpoint (EP) detection is a key component of far-field speech recognition systems that assist the user through voice commands. The endpoint detector has to trade-off between accuracy and latency, since waiting longer reduces the cases of users being cut-off early. We propose a novel two-pass solution for endpointing, where the utterance endpoint detected from a first pass endpointer is verified by a 2nd-pass model termed EP Arbitrator. Our method improves the trade-off between early cut-offs and latency over a baseline endpointer, as tested on datasets including voice-assistant transactional queries, conversational speech, and the public SLURP corpus. We demonstrate that our method shows improvements regardless of the first-pass EP model used.
Since humans can listen to audio and watch videos at faster speeds than actually observed, we often listen to or watch these pieces of content at higher playback speeds to increase the time efficiency of content comprehension. To further utilize this capability, systems that automatically adjust the playback speed according to the user's condition and the type of content to assist in more efficient comprehension of time-series content have been developed. However, there is still room for these systems to further extend human speed-listening ability by generating speech with playback speed optimized for even finer time units and providing it to humans. In this study, we determine whether humans can hear the optimized speech and propose a system that automatically adjusts playback speed at units as small as phonemes while ensuring speech intelligibility. The system uses the speech recognizer score as a proxy for how well a human can hear a certain unit of speech and maximizes the speech playback speed to the extent that a human can hear. This method can be used to produce fast but intelligible speech. In the evaluation experiment, we compared the speech played back at a constant fast speed and the flexibly speed-up speech generated by the proposed method in a blind test and confirmed that the proposed method produced speech that was easier to listen to.
To promote speech processing and recognition research in driving scenarios, we build on the success of the Intelligent Cockpit Speech Recognition Challenge (ICSRC) held at ISCSLP 2022 and launch the ICASSP 2024 In-Car Multi-Channel Automatic Speech Recognition (ICMC-ASR) Challenge. This challenge collects over 100 hours of multi-channel speech data recorded inside a new energy vehicle and 40 hours of noise for data augmentation. Two tracks, including automatic speech recognition (ASR) and automatic speech diarization and recognition (ASDR) are set up, using character error rate (CER) and concatenated minimum permutation character error rate (cpCER) as evaluation metrics, respectively. Overall, the ICMC-ASR Challenge attracts 98 participating teams and receives 53 valid results in both tracks. In the end, first-place team USTCiflytek achieves a CER of 13.16% in the ASR track and a cpCER of 21.48% in the ASDR track, showing an absolute improvement of 13.08% and 51.4% compared to our challenge baseline, respectively.
The emergence of industrial-scale speech recognition (ASR) models such as Whisper and USM, trained on 1M hours of weakly labelled and 12M hours of audio only proprietary data respectively, has led to a stronger need for large scale public ASR corpora and competitive open source pipelines. Unlike the said models, large language models are typically based on Transformer decoders, and it remains unclear if decoder-only models trained on public data alone can deliver competitive performance. In this work, we investigate factors such as choice of training datasets and modeling components necessary for obtaining the best performance using public English ASR corpora alone. Our Decoder-Only Transformer for ASR (DOTA) model comprehensively outperforms the encoder-decoder open source replication of Whisper (OWSM) on nearly all English ASR benchmarks and outperforms Whisper large-v3 on 7 out of 15 test sets. We release our codebase and model checkpoints under permissive license.
Multi-talker overlapped speech recognition remains a significant challenge, requiring not only speech recognition but also speaker diarization tasks to be addressed. In this paper, to better address these tasks, we first introduce speaker labels into an autoregressive transformer-based speech recognition model to support multi-speaker overlapped speech recognition. Then, to improve speaker diarization, we propose a novel speaker mask branch to detection the speech segments of individual speakers. With the proposed model, we can perform both speech recognition and speaker diarization tasks simultaneously using a single model. Experimental results on the LibriSpeech-based overlapped dataset demonstrate the effectiveness of the proposed method in both speech recognition and speaker diarization tasks, particularly enhancing the accuracy of speaker diarization in relatively complex multi-talker scenarios.
While massively multilingual speech models like wav2vec 2.0 XLSR-128 can be directly fine-tuned for automatic speech recognition (ASR), downstream performance can still be relatively poor on languages that are under-represented in the pre-training data. Continued pre-training on 70-200 hours of untranscribed speech in these languages can help -- but what about languages without that much recorded data? For such cases, we show that supplementing the target language with data from a similar, higher-resource 'donor' language can help. For example, continued pre-training on only 10 hours of low-resource Punjabi supplemented with 60 hours of donor Hindi is almost as good as continued pretraining on 70 hours of Punjabi. By contrast, sourcing data from less similar donors like Bengali does not improve ASR performance. To inform donor language selection, we propose a novel similarity metric based on the sequence distribution of induced acoustic units: the Acoustic Token Distribution Similarity (ATDS). Across a set of typologically different target languages (Punjabi, Galician, Iban, Setswana), we show that the ATDS between the target language and its candidate donors precisely predicts target language ASR performance.