Voice tech

Interspeech 2021: Take-aways on Automatic Speech Recognition

Janne Pylkkönen

Sep 28, 2021

4 min read

Our report from Interspeech, the largest scientific conference focusing on speech science and technology.

  • Copy link

  • Mail

  • LinkedIn

  • Facebook

  • Twitter

This year’s conference was held in the beginning of September in Brno, Czech Republic. Typically there would be some 2000 attendees at the conference, but due to Covid-19, this year most of the attendees joined the conference virtually. I was there on site with 350+ other researchers, and here are my impressions on the scientific catering in the field of automatic speech recognition (ASR).

New datasets

ASR is a data-heavy field. Industry leaders are using tens of thousands of hours of transcribed speech to train their models, but most of the ASR research has relied on much smaller publicly available corpora. Only very recently opportunities for using larger non-proprietary speech corpora have emerged. This year Facebook published Multilingual LibriSpeech (MLS), but that is limited to read-speech data. Now at Interspeech, two large ASR corpora were published, which extend the available domains:

Also worth checking is Facebook's research paper which showed that an ASR model trained on publicly available corpora, combined with fine-tuning to target data, works well for real-world tasks:


wav2vec is an unsupervised (or "self-supervised" as they like to call it) method for learning speech representation. Its latest incarnation, wav2vec 2.0, is gaining popularity: At Interspeech there were 10 papers mentioning wav2vec in the title, and many more which used it in their experiments. The benefit in using such pre-trained representations are a drastic drop in the training data requirements, thus making it attractive for limited resource scenarios.

A nice analysis on the nature of wav2vec 2.0 was provided by Facebook in their paper:

  • Robust wav2vec 2.0: Analyzing Domain Shift in Self-Supervised Pre-Training, Wei-Ning Hsu (Facebook, USA), Anuroop Sriram (Facebook, USA), Alexei Baevski (Facebook, USA), Tatiana Likhomanenko (Facebook, USA), Qiantong Xu (Facebook, USA), Vineel Pratap (Facebook, USA), Jacob Kahn (Facebook, USA), Ann Lee (Facebook, USA), Ronan Collobert (Facebook, USA), Gabriel Synnaeve (Facebook, France) and Michael Auli (Facebook, USA)

Trends in ASR

Rather than introducing a multitude of complex new network architectures, this year's focus appeared to be more on the practical side of ASR: Reducing the streaming latency, fitting the models on-device, and overall reducing computation and memory footprint. Various improvements were presented towards these goals, especially for transformer-transducer models.

For examples, see the following papers:

  • Improving Streaming Transformer Based ASR Under a Framework of Self-Supervised Learning, Songjun Cao (Tencent, China), Yueteng Kang (Tencent, China), Yanzhe Fu (Tencent, China), Xiaoshuo Xu (Tencent, China), Sining Sun (Tencent, China), Yike Zhang (Tencent, China) and Long Ma (Tencent, China)

  • An Efficient Streaming Non-Recurrent On-Device End-to-End Model with Improvements to Rare-Word Modeling, Tara N. Sainath (Google, USA), Yanzhang He (Google, USA), Arun Narayanan (Google, USA), Rami Botros (Google, USA), Ruoming Pang (Google, USA), David Rybach (Google, USA), Cyril Allauzen (Google, USA), Ehsan Variani (Google, USA), James Qin (Google, USA), Quoc-Nam Le-The (Google, USA), Shuo-Yiin Chang (Google, USA), Bo Li (Google, USA), Anmol Gulati (Google, USA), Jiahui Yu (Google, USA), Chung-Cheng Chiu (Google, USA), Diamantino Caseiro (Google, USA), Wei Li (Google, USA), Qiao Liang (Google, USA) and Pat Rondon (Google, USA)

  • Reducing Streaming ASR Model Delay with Self Alignment, Jaeyoung Kim (Google, USA), Han Lu (Google, USA), Anshuman Tripathi (Google, USA), Qian Zhang (Google, USA) and Hasim Sak (Google, USA)

Looking for something more exotic? Check out the research on non-autoregressive ASR:

RNN transducers are still going strong. Several publications had adopted the Hybrid Autoregressive Transducer (HAT) approach for combining external language models with the end-to-end model.

And more...

Interspeech is a lot more than just an ASR conference, too much to cover in a single blog post. One interesting and timely topic was the COVID-19 challenge: detecting infection based on cough and speech samples!

You can browse the full list of publications at

Latest blog posts

Voice tech

Create a WebRTC Video Chat App With Speechly Transcription

Learn how to build a WebRTC video chat application that uses the Speechly Browser Client to transcribe audio from a MediaStream.

Mathias Lindholm

May 24, 2022

8 min read

Voice design

The Fastest UI for the Web and Mobile

Abandoning the Voice Assistant model for a Voice UI as a Feature results in the most efficient UI since the Touchscreen.

Collin Borns

May 12, 2022

2 min read

Voice design

Evolution of UIs

From Punched Cards to Touch Screens, User Interfaces have evolved significantly. Will Voice be next in this evolution?

Collin Borns

May 11, 2022

4 min read