CoChat

Towards Better Decoding and Language Model Integration in Sequence to Sequence Models

Open access
Chorowski, Jan, Jaitly, Navdeep · 2017 · Interspeech 2017

The DOI matches its Crossref record; 47 of the 47 references it cites were found in a registry or an index; none of the 9 sources its claims cite is checked for support yet.

Abstract

The recently proposed Sequence-to-Sequence (seq2seq) framework advocates replacing complex data processing pipelines, such as an entire automatic speech recognition system, with a single neural network trained in an end-to-end fashion.In this contribution, we analyse an attention-based seq2seq speech recognition system that directly transcribes recordings into characters.We observe two shortcomings: overconfidence in its predictions and a tendency to produce incomplete transcriptions when language models are used.We propose practical solutions to both problems achieving competitive speaker independent word error rates on the Wall Street Journal dataset: without separate language models we reach 10.6% WER, while together with a trigram language model, we reach 6.7% WER.

Overview

CoChat Summary

TL;DR

This paper analyzes two failure modes of an attention-based seq2seq speech recognizer: overconfident character predictions and incomplete transcripts when using language models. It proposes label smoothing plus decode-time coverage handling, reaching 10.6% WER without a separate language model and 6.7% WER with a trigram language model on WSJ.

Key findings

  • The authors report that the baseline seq2seq model on WSJ achieved 14.2% WER on eval92, while unigram label smoothing reduced this to 10.6% and temporal/neighborhood smoothing to 10.7% on eval92.
  • With an extended trigram language model and coverage, the best reported result was 6.7% WER on eval92; the corresponding dev93 score was 9.7% for temporal label smoothing + coverage.
  • They found that model overconfidence makes beam search less effective and that increasing SoftMax temperature can help beam search, but without changing greedy decoding accuracy.
  • For incomplete transcripts, simply constraining EOS was not sufficient at large beam widths; promoting coverage helped, while maximizing transcript length alone caused more word insertions and worse overall WER.

Methods

  • Attention-based seq2seq speech recognition model built on Listen, Attend and Spell: a multilayer Bi-LSTM listener, an LSTM speller, and attention.
  • Experiments were run on the Wall Street Journal dataset, training on si284, validating on dev93, and evaluating on eval92; inputs were 80-dimensional mel-scale filterbanks with first and second differences and per-speaker normalization.
  • Main analysis compared decoding strategies and regularizers: SoftMax temperature changes, label smoothing variants (uniform/unigram and neighborhood/temporal), beam search settings, and a decode-time coverage penalty with an extended trigram language model.

Limitations

  • The paper states that adding a language model introduces two challenges: model overconfidence makes balancing the beam-search terms difficult, and incomplete transcripts occur unless a coverage term is added.
  • The authors note that simply maximizing transcript length yields more word insertion errors and worse overall WER.
  • Evident from the text: results are reported on a single dataset (WSJ) with a particular ASR setup, so generalization beyond this setting is not established in the paper.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.21437/interspeech.2017-343

Cites

47 of 47 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

0 Valid · 2 Not valid · 7 Not checked · 37 Own results and interpretations

20 more claims

Create a free account to see all 46 claims and to verify the 9 references still open.

See all claims — free account
  • It may state two separate claims as one. 16
  • Probably not a substantial claim. 5
  • The quote only partly states this claim. 2

Create a free account to see these claims.

Citations

Cites

Show all 47 references

Research works printed but not linked: 2

  1. [1][1] Y. LeCun, Y. Bengio, and G. Hinton, 'Deep learning,' Nature , vol. 521, no. 7553, pp. 436-444, 2015. -
  2. [23][23] A. Graves and N. Jaitly, 'Towards End-To-End Speech Recognition with Recurrent Neural Networks,' in ICML '14 , 2014, pp. 17641772. -

Cited by

Items 201–300 of 398

Metadata

Authors
Chorowski, J., & Jaitly, N.
Affiliations
Google (United States) (US)
Venue
Interspeech 2017
Publisher
ISCA
Published
2017-08-20
Pages
523-527
ISSN
2308-457X (ISSN-L)
OpenAlex
W2577366047
Cited by
358
References
34
Citations by year
2026: 4 · 2025: 14 · 2024: 13 · 2023: 39 · 2022: 26 · 2021: 73 · 2020: 88 · 2019: 49 · 2018: 40 · 2017: 12
FWCI
23.58
Type
conference paper
Language
en
Open access
Locations
  • Interspeech (USB) (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Speech Recognition and Synthesis
  • Physical Sciences › Computer Science › Signal Processing › Speech and Audio Processing
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Handwritten Text Recognition Techniques
Keywords
attention-based models, language model integration, trigram language model, end-to-end speech recognition, speaker-independent recognition, Wall Street Journal

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-10-11, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-11 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.