News

On the theme of invariance, we would also like to announce an upcoming Special Session at Interspeech 2023 (not directly linked to ZeroSpeech) which will likely be of interest to members of the community. The theme is “Invariant and Robust Pretrained Acoustic Models”:

perceptimatic.github.io/irpam2023

The session seeks to kick off an important discussion about questions about invariances and robustness in pre-trained acoustic models. Do the representations remain consistent across instances of the same phoneme in different phonetic contexts (i.e., are they phonemic or merely allophone representations)? Do they hold up under noise and distortions? Are they invariant to different talkers and/or accents? The 2023 Interspeech Special Session on Invariant and Robust Pretrained Acoustic Models (IRPAM) aims to address both the evaluation problem and the problem of invariance in pretrained acoustic models.

Submissions for IRPAM will follow the same schedule and procedure as the main Interspeech 2023 conference, which will be held in-person in Dublin interspeech2023.org

Paper submission deadline: March 1st, 2023, 23:59, Anywhere on Earth.

Paper update deadline: March 8th, 2023, 23:59, Anywhere on Earth.

Author notification: May 17th, 2023

Interspeech conference dates: August 20th to 24th, Dublin, Ireland

Please see the IRPAM 2023 website for more details.

Zero Resource Speech Benchmark

As the ZeroSpeech organizers, we are pleased to announce that ZeroSpeech (https://zerospeech.com/) has undergone some long-awaited updates which are now available for general usage. We have :

• completely redone the website
• made it easier to download the datasets and evaluate the results
• created a command-line utility to do evaluation locally, and upload the results/embeddings so that scores appear on the dynamic leaderboard
• created an entirely new in-house system for submitting models (to be open publicly in February 2023), which replaces the old submission system that often posed technical difficulties
• reorganized the evaluation into four permanent Benchmarks (Task 1, 2, 3 and 4), which are no longer tied to specific challenge events.

Submissions thus are now open all year round! Tasks 1 (acoustic modelling), 2 (spoken term discovery), and 4 (language modelling) are currently open for submissions. We would also like to draw your attention towards a new set of conditions in ABX-LS benchmarks that propose to measure phonemic context-invariance in learned representations. Details are presented on the leaderboard (zerospeech.com/tasks/task_1/results/) and in this paper (arxiv.org/abs/2210.15775).

Our zerospeech.com website has undergone some well-awaited updates

• Refactoring of all the benchmarks and leaderboards
  • Task 1: Acoustic Unit Discovery / Speech Representation Learning
  • Task 2: Spoken term Discovery / Word segmentation
  • Task 3: Discrete Resynthesis / TTS without T
  • Task 4: Spoken Language Modeling
• Republishing datasets containing gold data to allow local evaluation
• Submissions are now open forever
• New toolkit to help with evaluation/download/upload data
• New upload functionality with local zerospeech accounts, no need to use codalab anymore (Will be available in February 2023).

The NeurIPS Schedule is now finalized, and the 2021 Enhanced Zero Resource Speech Challenge will be featured during the Competition Track on Day 4, December 10th, 2021. The winners will be announced at this time. Here is the detailed schedule. All times are UTC.

• Main competition track plenary session: Day 4 Competition Track Session held from 10:00 AM to 12:05 PM UTC. The Zero Speech plenary will be from 10:25 to 10:45 AM UTC.

• Breakout session (Zoom): The Zero Speech competition breakout session will be held at 12:30 PM UTC, and will consist of short system overviews and an open discussion looking forward to the future of the Zero Resource challenges/evaluations. We look forward to seeing you there!

Please note that conference rules state that participants in either the plenary session or the breakout session must be registered to NeurIPS. Zoom link will be made available via the NeurIPS Conference Platform.

We invite ZeroSpeech 2021 participants who are submitting their systems to the NeurIPS Competition to submit non-archival papers to the 2nd Self-supervised Learning for Audio and Speech Processing @ AAAI 2022. Note that participants wishing to submit papers to the workshop are encouraged to get their test results early by submitting their systems no later than November 3rd, 2021, as the workshop paper deadline is nine days later. Systems can still be submitted up until November 19th for inclusion in the NeurIPS Competition. Participants are also encouraged to submit an (archival) paper to the IEEE JSTSP Special Issue (see below.)

Please see the workshop website for more information.

• Website: https://aaai-sas-2022.github.io/
• Workshop Date: Between February 28th and March 1st, 2022
• Submission Deadline: November 12th, 2021 (Anywhere on Earth, AoE)
• Submission website: https://cmt3.research.microsoft.com/SAS2022

As previously mentioned, a workshop on Semi-Supervised Representations for Speech and Audio Processing will be held at AAAI in Vancouver, in February 2022. A session will be reserved for discussion of ZeroSpeech, and we encourage participants to submit (non-archival) workshop papers. The CFP will be posted here soon.

In order to accomodate the tight AAAI deadlines, we will encourage early submissions, submitted by November 3rd at the latest for those wish to include systems in AAAI workshop papers. The workshop paper deadline is November 12th. While the evaluation is automated, this buffer time will allow you to receive your results with ample margin.

The deadline to have systems submitted in time to be considered for the NeurIPS Competition remains November 19th, and teams making early submissions may of course submit additional systems later on.

Finally, we remind participants of the IEEE JSTSP Special Issue on Self-Supervised Learning for Speech and Audio Processing. Submissions to both the workshop and the special issue are possible. Please contact us for more information if you are planning to submit to both.

Interspeech 2021 is fast approaching, and we encourage all registered participants to please check out the Interspeech submissions to the Challenge, which represent the first round of evaluation of the 2021 Challenge on Spoken Language Modelling. Three-minute preview videos and full papers are already up on the Interspeech conference platform (https://interspeech2021.gcon.me/). To see our presenters live, please note the following session information:

• Tuesday, August 31, 19:00 CEST, in-person in Brno: E-O-1-1, Information Retrieval for ZeroSpeech 2021: The Submission by University of Wroclaw
• Wednesday, September 1, 11:00 CEST, online poster session: M-V-2-6, Analyzing Speaker Information in Self-Supervised Models to Improve Zero-Resource Speech Processing
• Wednesday, September 1, 11:00 CEST, online poster session: M-V-2-8, Speech Representation Learning Combining Conformer CPC with Deep Cluster for the ZeroSpeech Challenge 2021
• Wednesday, September 1, 11:00 CEST, online poster session: M-V-2-10, The Zero Resource Speech Challenge 2021: Spoken Language Modelling

Please come and check out the great work!

As previously announced, the second round of submissions to ZeroSpeech, including the multi-modal track, are an official NeurIPS competition. The competition will have a pre-recorded overview talk by the organizers, presenting the overview of the Enhanced ZeroSpeech 2021 Challenge and its evaluations. The competition will also have a live, virtual session in which submitted systems, results, and winners will be presented. Watch this space for more details about the live session.

The submission deadline for systems to be included in the NeurIPS competition is November 19th, 2021, as previously announced. This will give us time to run the evaluation and prepare the material for the live session. If there is sufficient interest from participants, information about early submissions (made before October 20th) may also be included in the pre-recorded overview talk.

We invite participants to submit papers describing their system and results for the IEEE JSTSP Special Issue on Self-Supervised Learning for Speech and Audio Processing. This special issue has a submission deadline of December 31, 2021. Please see the call for papers for the IEEE JSTSP special issue for detailed information.

Please be advised that the previously proposed workshop on semi-supervised representations will not take place at NeurIPS this year. The workshop has been re-submitted to AAAI (Jan/Feb 2022), and, if accepted ZeroSpeech 2021 will have a presence. Watch this space for details.

As previously announced, we are opening the 2021 benchmark on unsupervised spoken language modelling to a second round of submissions as a NeurIPS competition.

The formulation for this enhanced edition of the challenge is slightly different. While the benchmark remains the same, submissions will be divided into two tracks.

:ref:Track 1 <ZRC2021_2021n_track_s> corresponds to the classic, purely unsupervised case, while :ref:Track 2 <ZRC2021_2021n_track_m> allows for weak, indirect supervision in the form of images or videos. See the detail pages for more information.

We continue to solicit submissions of models of all sizes. We strongly encourage the exploration of large language models at NeurIPS 2021, in order to help better understand the problem space.

Individual research groups may make no more than four submissions each.

The competition will be paired with a planned NeurIPS workshop on unsupervised and semi-supervised speech modelling. Watch this space for more details.

We are pleased to announce that the Zero Resource Speech Challenge Enhanced Edition 2021 has been accepted as a NeurIPS competition. Details, including dates and submission procedures, will be announced in early May 2021.

We are pleased to announce that the Zero Resource Speech Challenge 2021 has been accepted as a special session at INTERSPEECH (interspeech2021.org). We look forward to seeing you!

• December 3rd, 2021: Decisions for AAAI-SAS-2022 workshop papers
• December 10th, 2021: NeurIPS competition track session
• December 31, 2021: Paper submission deadline for IEEE JSTSP Special Issue on Self-Supervised Learning for Speech and Audio Processing.

We released the BERT and LSTM baselines on our github. The baselines submissions are also available at download.zerospeech.com.

We are pleased to announce the Zero Resource Speech Challenge 2021 aiming at Spoken Language Modeling. We released challenge material (datasets, evaluation software and submission procedure), please see the Tasks and intended goal and the Instruction pages for details.

Following the decision of the Interspeech organizers to shift the dates of the conference to 26-29 October 2020, we are shifting the ZeroSpeech Challenge deadline to 24 April 2020. See the updated timeline for details.

• The codalab competition is open: https://competitions.codalab.org/competitions/22999.

• The baseline is available for download at https://download.zerospeech.com.

• We released an updated version of the evaluation software. It simplifies the command line options and the output is now formatted as JSON. See Evaluation.

• To update the software, please remove the current virtual environment (conda deactivate && conda env remove -n zerospeech2020) and reinstall the software from scratch by following the instructions here: https://github.com/bootphon/zerospeech2020.

We have updated the 2020 challenge website for general information as well as more detailed information, including information about downloading the updated software tools and datasets.

We are happy to inform you that the ZeroSpeech 2020 Challenge has been accepted as a special session at Interspeech2020. This edition is a consolidating challenge in which participants submit systems to the Zero Resource 2017 (track 1 or track 2) or the Zero Resource 2019 tasks

We are pleased to announce the Zero Resource Speech Challenge 2020, which has been submitted as a special session at Interspeech 2020. ZeroSpeech 2020 is a consolidating challenge in which participants submit systems to the Zero Resource 2017 (Track 1 or Track 2) or the Zero Resource 2019 tasks. Participants are particularly encouraged to submit to multiple tracks/challenges (unit discovery evaluated on both the 2017 and 2019 evaluations, unit discovery used as a basis for spoken term discovery).

• Training data. The training data available will be exactly the same as in the 2017 and 2019 challenges, separated into the same development and surprise languages. Participants interested in starting work immediately are encouraged to begin by obtaining the existing development kits and datasets. Please see ZeroSpeech 2017 and ZeroSpeech 2019: TTS without T to get started.

• Test data and evaluation. The test data and evaluations will undergo some minor modifications (such as bug fixes). Updated kits will be available in early 2020.

• Sign-up, submission procedure, and leaderboard. The challenge will again be run again on Codalab, with a leaderboard allowing for comparison between 2017/2019 evaluations for combined submissions.

The schedule for the ZeroSpeech special session at Interspeech 2019 is now available here. We are looking forward to seeing you there!

The surprise language for the Interspeech 2019 Zero Ressource Speech Challenge is now available for download. To get the files, three simple steps:

1. register in Codalab here

2. download the data here

3. update your evaluation package in the docker image of the Development kit:

git pull
source activate eval
pip install pyyaml

Please remember the next deadline is for wavefile submission in Codalab on March 15, 2019, 23h59 GMT - 12. This deadline will be strictly enforced.

We are happy to inform you that the ZeroSpeech 2019 Challenge has been accepted as a special session. If you submit a paper for this challenge, it will be reviewed in conjunction with all other papers submitted to Interspeech 2019.