On lattice generation for large vocabulary speech recognition

Rybach, David; Riley, Michael; Schalkwyk, Johan

doi:10.1109/asru.2017.8268940

Cited by 8 publications

(4 citation statements)

References 27 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…All previous biasing work, for both conventional and E2E models [2,4,6], combines scores from the contextual LM and the base model (ASR LM or E2E model) on the word or subword lattice, a concept known as on-the-fly lattice rescoring [8]. E2E models are decoded with tight beam thresholds resulting in a much smaller number of path hypotheses compared to conventional models.…”

Section: Biasing Before Beam Pruningmentioning

confidence: 99%

Shallow-Fusion End-to-End Contextual Biasing

Zhao¹,

Sainath²,

Rybach³

et al. 2019

Interspeech 2019

Self Cite

100

View full text Add to dashboard Cite

Contextual biasing to a specific domain, including a user's song names, app names and contact names, is an important component of any production-level automatic speech recognition (ASR) system. Contextual biasing is particularly challenging in end-toend models because these models keep a small list of candidates during beam search, and also do poorly on proper nouns, which is the main source of biasing phrases. In this paper, we present various algorithmic and training improvements to shallow-fusionbased biasing for end-to-end models. We will show that the proposed approach obtains better performance than a state-ofthe-art conventional model across a variety of tasks, the first time this has been demonstrated.

show abstract

Section: Biasing Before Beam Pruningmentioning

confidence: 99%

Shallow-Fusion End-to-End Contextual Biasing

Zhao¹,

Sainath²,

Rybach³

et al. 2019

Interspeech 2019

Self Cite

100

View full text Add to dashboard Cite

show abstract

“…These results motivate future work to improve lattice generation [18,19], particularly in E2E ASR systems. Our current research also explores open-vocabulary decoding in a WFST framework, in which novel words may be included in a lattice and derived phrase alternatives.…”

Section: Discussionmentioning

confidence: 70%

Toward Zero Oracle Word Error Rate on the Switchboard Benchmark

Faria¹,

Janin²,

Riedhammer³

et al. 2022

Preprint

View full text Add to dashboard Cite

The "Switchboard benchmark" is a very well-known test set in automatic speech recognition (ASR) research, establishing record-setting performance for systems that claim human-level transcription accuracy. This work highlights lesser-known practical considerations of this evaluation, demonstrating major improvements in word error rate (WER) by correcting the reference transcriptions and deviating from the official scoring methodology. In this more detailed and reproducible scheme, even commercial ASR systems can score below 5% WER and the established record for a research system is lowered to 2.3%. An alternative metric of transcript precision is proposed, which does not penalize deletions and appears to be more discriminating for human vs. machine performance. While commercial ASR systems are still below this threshold, a research system is shown to clearly surpass the accuracy of commercial human speech recognition. This work also explores using standardized scoring tools to compute oracle WER by selecting the best among a list of alternatives. A phrase alternatives representation is compared to utterance-level N-best lists and word-level data structures; using dense lattices and adding out-of-vocabulary words, this achieves an oracle WER of 0.18%.

show abstract

“…Using a limited beam size H, the search generates an N -best list with a computational cost of O(H). However, N -best lists are insufficient for some downstream applications or post-processing stages, such as (1) language model rescoring [10][11][12]; (2) downstream processing of ASR output (e.g., translation [13] and keyword spotting [14,15]); (3) confusion network generation [16,17]; and (4) sequence discriminative training [18][19][20].…”

Section: Introductionmentioning

confidence: 99%

“…Hypothesis merging (or path recombination) is the main difference between regular beam search and lattice-based search. For conventional methods, Rybach et al divided merging procedures into two categories [12]: the phone-pair approach [10,21] and the N -best history approach [22][23][24]. However, neither are suitable for end-to-end models because they require additional n-gram language models and HMM-based acoustic models to generate lattices.…”

Section: Introductionmentioning

confidence: 99%

VQ-T: RNN Transducers using Vector-Quantized Prediction Network States

Shi¹,

Saon²,

Haws³

et al. 2022

Interspeech 2022

View full text Add to dashboard Cite

Beam search, which is the dominant ASR decoding algorithm for end-to-end models, generates tree-structured hypotheses. However, recent studies have shown that decoding with hypothesis merging can achieve a more efficient search with comparable or better performance. But, the full context in recurrent networks is not compatible with hypothesis merging. We propose to use vector-quantized long short-term memory units (VQ-LSTM) in the prediction network of RNN transducers. By training the discrete representation jointly with the ASR network, hypotheses can be actively merged for lattice generation. Our experiments on the Switchboard corpus show that the proposed VQ RNN transducers improve ASR performance over transducers with regular prediction networks while also producing denser lattices with a very low oracle word error rate (WER) for the same beam size. Additional language model rescoring experiments also demonstrate the effectiveness of the proposed lattice generation scheme.

show abstract

On lattice generation for large vocabulary speech recognition

Cited by 8 publications

References 27 publications

Shallow-Fusion End-to-End Contextual Biasing

Shallow-Fusion End-to-End Contextual Biasing

Toward Zero Oracle Word Error Rate on the Switchboard Benchmark

VQ-T: RNN Transducers using Vector-Quantized Prediction Network States

Contact Info

Product

Resources

About