TED-LIUM 3: twice as much data and corpus repartition for experiments on speaker adaptation

By François Hernandez, Sahar Ghannay, Natalia A. Tomashenko, Yannick Estève. 2018

434 citations recorded by the source corpus.

In this paper, we present TED-LIUM release 3 corpus (TED-LIUM 3 is available on https://lium.univ-lemans.fr/ted-lium3/) dedicated to speech recognition in English, which multiplies the available data to train acoustic models in comparison with TED-LIUM 2, by a factor of more than two. We present the recent development on Automatic Speech Recognition (ASR) systems in comparison with the two previous releases of the TED-LIUM Corpus from 2012 and 2014. We demonstrate that, passing from 207 to 452 h of transcribed speech training data is really more useful for end-to-end ASR systems than for HMM-based state-of-the-art ones. This is the case even if the HMM-based ASR system still outperforms the end-to-end ASR system when the size of audio training data is 452 h, with a Word Error Rate (WER) of 6.7% and 13.7%, respectively. Finally, we propose two repartitions of the TED-LIUM release 3 corpus: the legacy repartition that is the same as that existing in release 2, and a new repartition, calibrated and designed to make experiments on speaker adaptation. Similar to the two first releases, TED-LIUM 3 corpus will be freely available for the research community.

Subfields: Artificial Intelligence, Physiology, Evaluation & Benchmarks, Speech & Audio, Signal Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Physiology

Ryuichi Yamamoto

In Evaluation & Benchmarks

In Speech & Audio

In Signal Processing

Citations in the corpus

Robust Speech Recognition via Large-Scale Weak Supervision (2022, methodology): “These include a long-form adaptation of TED-LIUM3 (Hernandez et al., 2018) concatenated so that each example is a full-length TED talk, a collection of jargon-laden segments taken from The Late Show with Stephen Colbert (Meanwhile), sets of videos/podcasts that has been used as ASR benchmarks in onl”
BigSSL: Exploring the Frontier of Large-Scale Semi-Supervised Learning for Automatic Speech Recognition (2021, methodology): “Data: The SpeechStew [30] dataset is assembled by putting together seven public speech corpora—AMI [74], Common Voice [75], English Broadcast News,2 LibriSpeech [28], Switchboard/Fisher,3 TED-LIUM v3 [76], [77] and Wall Street Journal.”
GigaSpeech: An Evolving, Multi-domain ASR Corpus with 10, 000 Hours of Transcribed Audio (2021, background): “It consists of 452 hours of audio from TED talks in its latest release TED-LIUM 3.”
Robust wav2vec 2.0: Analyzing Domain Shift in Self-Supervised Pre-Training (2021, methodology): “The six domains are: (1) LibriSpeech [26] (LS) and Libri-light [27] (LL), which contain about 960 and 60K hours of 16kHz crowd-sourced audiobook recordings, respectively, derived from the LibriVox project; (2) TED-LIUM v3 [28] (TD), which contains 452 hours of TED conference audio in 16kHz; (3) Swit”
Rethinking Evaluation in ASR: Are Our Models Robust Enough? (2020, methodology): “TED-LIUM v3 (TL) [19] is based on TED conference videos.”
A Comparative Study on Transformer vs RNN in Speech Applications (2019, background): “AISHELL [28] zh 170 read dev / test AURORA4 [29] (*) en 15 noisy read (dev 0330) A / B / C / D CSJ [30] ja 581 spontaneous eval1 / eval2 / eval3 CHiME4 [31] (*) en 108 noisy far-field multi-ch read dt05 simu / dt05 real / et05 simu / et05 real CHiME5 [32] en 40 noisy far-field multi-ch conversationa”
Unsupervised Domain Adaptation for Speech Recognition via Uncertainty Driven Self-Training (2020, methodology): “Data: We use the Wall Street Journal [26] (WSJ) dataset as our source domain, and TED-LIUM 3 [27] (TED) as well as Switchboard [28] (SWBD) as our target domains.”
UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled Data (2021, methodology): “We use our UniSpeech model to pre-train phoneme representations in reading English domain, namely on Librispeech, and transfer them to spoken English domain on Tedlium3 (Hernandez et al., 2018) dataset.”
Very Deep Self-Attention Networks for End-to-End Speech Recognition (2019, methodology): “nge dropout values for this model, so each layer’s hidden layers are dropped more severely Table 3: The Transformer results on the TED-LIUM test set using TED-LIUM 3 training set. Models Test WER CTC [19] 17.4 CTC/LM + speed perturbation [19] 13.7 12Enc-12Dec (Ours) 14.2 Stc. 12Enc-12Dec (Ours) 12.4”
Performance-Efficiency Trade-Offs in Unsupervised Pre-Training for Speech Recognition (2021, methodology): “SEW-D outperforms W2V2 in most pre-training settings, when experimenting with LibriSpeech (Panayotov et al., 2015), Ted-lium 3 (Hernandez et al., 2018), VoxPopuli (Wang et al., 2021a), and Switchboard (Godfrey and Holliman, 1993) datasets.”
FINDINGS OF THE IWSLT 2020 EVALUATION CAMPAIGN (2020, neutral): “…• How2 (Sanabria et al., 2018)8 • LibriVoxDeEn (Beilharz and Sun, 2019)9 • Europarl-ST (Iranzo-Sánchez et al., 2020) • TED LIUM v2 (Rousseau et al., 2014) and v3 (Hernandez et al., 2018) • all the data provided by WMT 201910 • OpenSubtitles 2018 (Lison et al., 2018) • Augmented LibriSpeech…”
Google USM: Scaling Automatic Speech Recognition Beyond 100 Languages (2023, methodology): “The SpeechStew [2] dataset is assembled by putting together seven public speech corpora—AMI [78], Common Voice [79], English Broadcast News3, LibriSpeech [80], Switchboard/Fisher4, TED-LIUM v3 [81, 82] and Wall Street Journal5, which are all standard benchmarks [83–85] covering different domains in ”
SPGISpeech: 5, 000 hours of transcribed financial audio for fully formatted end-to-end speech recognition (2021, background): “Other corpora include TED-LIUM (450 hours of transcribed TED talks) [17] and Common Voice (a multilingual corpus of narrated prompts with ∼ 1,100 validated hours in English) [18], and GigaSpeech (10,000 hours from audiobooks, podcasts and YouTube videos) [19].”
CAPE: Encoding Relative Positions with Continuous Augmented Positional Embeddings (2021, methodology): “5h of training data, and TED-LIUM v3 (TL) [24], oratory speech with 452h of training data.”
Self-Supervised Speech Representation Learning: A Review (2022, positive): “Several datasets, including GigaSpeech [220], TED-LIUM 3 (TED3) [221], TED-LIUM 2 (TED2) [222], Switchboard (SWB) [223], TIMIT [96], and VoxLingua107 [224], are labeled”
SpeechBrain: A General-Purpose Speech Toolkit (2021, background): “Finally, our goal is to add support for additional languages and further expand the set of recipes to open-source datasets not yet available in the toolkit (e.g., TED-LIUM [101]).”