From 652b76424a8765eff7d47ad7a68923fa611268d5 Mon Sep 17 00:00:00 2001 From: Xuankai Chang Date: Fri, 19 Jan 2024 23:32:33 -0500 Subject: [PATCH 1/6] add initial version for interspeech2024 discrete speech unit challenge --- ...eech2024-Discrete-Speech-Unit-Challenge.md | 75 +++++++++++++++++++ 1 file changed, 75 insertions(+) create mode 100644 _posts/2024-01-19-Interspeech2024-Discrete-Speech-Unit-Challenge.md diff --git a/_posts/2024-01-19-Interspeech2024-Discrete-Speech-Unit-Challenge.md b/_posts/2024-01-19-Interspeech2024-Discrete-Speech-Unit-Challenge.md new file mode 100644 index 00000000..63885cbd --- /dev/null +++ b/_posts/2024-01-19-Interspeech2024-Discrete-Speech-Unit-Challenge.md @@ -0,0 +1,75 @@ +--- +layout: post +title: Interspeech2024 Speech Processing Using Discrete Speech Unit Challenge +description: This is the Interspeech2024 challenge website for speech processing using discrete speech unit challenge +date: 2024-01-19 09:00:00-0800 +comments: false +--- + + + +## Introduction + +In conventional speech processing approaches, models typically take either raw waveforms or high-dimensional features derived from these waveforms as input. For instance, spectral speech features continue to be widely employed, while learning-based deep neural network features have gained prominence in recent years. A promising alternative arises in the form of discrete speech representation, where speech signals within a temporal window can be represented by a single index $d$ within a range $K$: $d \in \{1, 2, \dots, K\}$. + +Three challenging tasks are proposed for using discrete speech representations. + 1. Automatic speech recognition (ASR): We will evaluate the ASR performance of the proposed systems on the proposed data. + 2. Text-to-speech (TTS): We will evaluate the quality of the generated speech. + 3. Singing voice synthesis (SVS): We will evaluate the quality of the synthesized singing voice. + + +Participation is open to all. Each team can participate in any task. This challenge has preliminarily been accepted as a special session for Interspeech 2024, and participants are strongly encouraged to submit papers to the special session. The focus of the special session is to promote the adoption of discrete speech representations and encourages novel insights. + + + +## Resources + + +### Baseline systems & toolkits +- [Automatic speech recognition (ASR)](https://github.com/espnet/espnet/tree/master/egs2) +- [Text-to-speech (TTS)]() +- [Singing voice synthesis (SVS)]() + +### Rules +* For each task, the training data must follow the baseline systems. However, there is no constraint on the data used in the foundation models. +* For submission, more details will be provided later. + + + +## Paper submission + +Papers for Interspeech Special Session have to be submitted following the same schedule and procedure as regular papers of [INTERSPEECH 2024](https://interspeech2024.org/). The submitted papers will undergo the same review process by anonymous and independent reviewers. + +Submission URL : (TBA) + + + +## Important dates +The schedule for the challenge is as follows +* Feb 20, 2024: Leaderboard is online and accepting submissions +* Mar 2, 2024: Paper Submission Deadline +* Mar 9, 2024: Paper Revision Deadline +* After Mar. 9: Leaderboard will still be open and new submissions will be evaluated + + + +## Organizers + +* Xuankai Chang (Carnegie Mellon University, U.S.) +* Jiatong Shi (Carnegie Mellon University, U.S.) +* Shinji Watanabe (Carnegie Mellon University, U.S.) +* Yossi Adi (Hebrew University, Israel) +* Xie Chen (Shanghai Jiao Tong University, China) +* Qin Jin (Renmin University of China, China) \ No newline at end of file From cb587a1b3f823fd44ebfd752d9218ce5722399ed Mon Sep 17 00:00:00 2001 From: ftshijt Date: Sat, 20 Jan 2024 17:58:42 -0500 Subject: [PATCH 2/6] update details in the challenge --- ...terspeech2024-Discrete-Speech-Unit-Challenge.md | 14 +++++++++++--- _site/index.html | 2 +- _site/news/announcement_1/index.html | 2 +- _site/projects/1_project/index.html | 2 +- _site/projects/2_project/index.html | 2 +- _site/projects/3_project/index.html | 2 +- _site/projects/4_project/index.html | 2 +- _site/projects/5_project/index.html | 2 +- _site/projects/6_project/index.html | 2 +- _site/publications/index.html | 2 +- 10 files changed, 20 insertions(+), 12 deletions(-) diff --git a/_posts/2024-01-19-Interspeech2024-Discrete-Speech-Unit-Challenge.md b/_posts/2024-01-19-Interspeech2024-Discrete-Speech-Unit-Challenge.md index 63885cbd..c923fd5c 100644 --- a/_posts/2024-01-19-Interspeech2024-Discrete-Speech-Unit-Challenge.md +++ b/_posts/2024-01-19-Interspeech2024-Discrete-Speech-Unit-Challenge.md @@ -33,12 +33,20 @@ Participation is open to all. Each team can participate in any task. This challe ### Baseline systems & toolkits - [Automatic speech recognition (ASR)](https://github.com/espnet/espnet/tree/master/egs2) -- [Text-to-speech (TTS)]() -- [Singing voice synthesis (SVS)]() +- [Text-to-speech (TTS)](https://github.com/espnet/espnet/tree/tts2/egs2/ljspeech/tts2) +- [Singing voice synthesis (SVS)](https://github.com/A-Quarter-Mile/espnet/tree/tmp_muskit/egs2/opencpop/svs2) +- [Discrete vocoder training](https://github.com/kan-bayashi/ParallelWaveGAN) + +### Dataset + +- ASR: [Librispeech](https://www.openslr.org/12) and [ML-SUPERB](https://drive.google.com/file/d/1zslKQwadZaYWXAmfBCvlos9BVQ9k6PHT/view?usp=sharing) +- TTS: [LJSpeech](https://keithito.com/LJ-Speech-Dataset/) and [Expresso](https://speechbot.github.io/expresso/) +- SVS: [Opencpop](https://wenet.org.cn/opencpop/) + ### Rules * For each task, the training data must follow the baseline systems. However, there is no constraint on the data used in the foundation models. -* For submission, more details will be provided later. +* For submission, more details will be provided later for each task. - +## Detailed tracks and rules + +### ASR Challenge + +* Data +* Framework +* Evaluation metrics +* Submission + +### TTS Challenge + +* Data +* Framework +* Evaluation metrics +* Submission + +### SVS Challenge + +* Data +* Framework +* Evaluation metrics +* Submission + +### Research in discrete representation of speech and audio + +* Call for research papers: As a special session, the track also accepts research papers in discrete representation of speech and audio. The paper could be related to any of the following topics: + * Discrete speech/audio/music representation learning + * Discrete representation application for any speech/audio processing downstream tasks (ASR, TTS, etc.) + * Evaluation of speech/audio discrete representation + * Efficient discrete speech/audio discrete representation + * Interpretability in discrete speech/audio discrete representation + * Other novel usage of discrete representation in speech/audio ## Paper submission From ff9a5492ef013a7483f69e262a27e1ad6bd3670f Mon Sep 17 00:00:00 2001 From: ftshijt Date: Sat, 20 Jan 2024 23:44:45 -0500 Subject: [PATCH 5/6] update the info --- ...eech2024-Discrete-Speech-Unit-Challenge.md | 56 ++++++++++++++++--- 1 file changed, 49 insertions(+), 7 deletions(-) diff --git a/_posts/2024-01-19-Interspeech2024-Discrete-Speech-Unit-Challenge.md b/_posts/2024-01-19-Interspeech2024-Discrete-Speech-Unit-Challenge.md index 6da33845..652bf39d 100644 --- a/_posts/2024-01-19-Interspeech2024-Discrete-Speech-Unit-Challenge.md +++ b/_posts/2024-01-19-Interspeech2024-Discrete-Speech-Unit-Challenge.md @@ -58,19 +58,53 @@ Participation is open to all. Each team can participate in any task. This challe * Evaluation metrics * Submission -### TTS Challenge +### TTS Challenge - Acoustic+Vocoder -* Data -* Framework -* Evaluation metrics +* Data: LJSpeech, following the train-dev-test split in [here](https://github.com/ftshijt/Interspeech2024_DiscreteSpeechChallenge). +* Framework: No framework restriction in TTS-Acoustic+Vocoder challenge, but the organizers have prepared the baseline training scripts (baseline model to be released soon) in [ESPnet](https://github.com/espnet/espnet/tree/tts2/egs2/ljspeech/tts2). +* Evaluation metrics: Mean cepstral distortion, F0 root mean square error, Bitrate, [UTMOS](https://github.com/sarulab-speech/UTMOS22/tree/master) +* Submission + * Submission package details: + * The synthesized voice of LJSpeech test set using full training set (with at least 16kHz). + * The synthesized voice of LJSpeech test set using 1-hour training set (with at least 16kHz). + * The discrete representation corresponding to the test set + * A technical report in Interspeech2024 paper format (no length limit) + * Notes: + * We encourage participants to additional contrast results in addition to their primary submission. However, due to the budget, we cannot gaurantee those all submitted results will be evaluated in the subjective metric. + * More details about the submission process will be updated later. + + +### TTS Challenge - Vocoder + +* Data: Expresso, following the train-dev-test split in [here](https://github.com/ftshijt/Interspeech2024_DiscreteSpeechChallenge) (Note that this is different from the original train-dev-test split in the benchmark paper). +* Framework: No framework restriction in TTS-Vocoder challenge, but the organizers have prepared the baseline training scripts (baseline model to be released soon) in [ESPnet](https://github.com/espnet/espnet/tree/tts2/egs2/ljspeech/tts2) and [ParallelWaveGAN](https://github.com/kan-bayashi/ParallelWaveGAN). +* Evaluation metrics: Mean cepstral distortion, F0 root mean square error, Bitrate, [UTMOS](https://github.com/sarulab-speech/UTMOS22/tree/master) * Submission + * Submission package details: + * The synthesized voice of LJSpeech test set using full training set (with at least 16kHz). + * The discrete representation corresponding to the test set + * A technical report in Interspeech2024 paper format (no length limit) + * Notes: + * We encourage participants to additional contrast results in addition to their primary submission. However, due to the budget, we cannot gaurantee those all submitted results will be evaluated in the subjective metric. + * More details about the submission process will be updated later. + ### SVS Challenge -* Data -* Framework +* Data: Opencpop, following the original segmentation and train/test split. +* Framework: No framework restriction in SVS challenge, but the organizers have prepared the baseline training scripts (baseline model to be released soon) in [ESPnet-Muskits](https://github.com/A-Quarter-Mile/espnet/tree/tmp_muskit/egs2/opencpop/svs2). * Evaluation metrics + * Objective metrics: Mean cepstral distortion, F0 root mean square error, Bitrate for efficiency measure + * Subjective metrics: Mean Opinion Score by organizers * Submission + * Submission package details: + * The synthesized voice of Opencpop test set (with at least 16kHz) + * The discrete representation corresponding to the test set + * A technical report in Interspeech2024 paper format (no length limit) + * Notes: + * We encourage participants to additional contrast results in addition to their primary submission. However, due to the budget, we cannot gaurantee those all submitted results will be evaluated in the subjective metric. + * More details about the submission process will be updated later. + ### Research in discrete representation of speech and audio @@ -81,6 +115,8 @@ Participation is open to all. Each team can participate in any task. This challe * Efficient discrete speech/audio discrete representation * Interpretability in discrete speech/audio discrete representation * Other novel usage of discrete representation in speech/audio +* Please refer to the "Paper submission" section for detail guidance of paper submission. + ## Paper submission @@ -110,4 +146,10 @@ The schedule for the challenge is as follows * Shinji Watanabe (Carnegie Mellon University, U.S.) * Yossi Adi (Hebrew University, Israel) * Xie Chen (Shanghai Jiao Tong University, China) -* Qin Jin (Renmin University of China, China) \ No newline at end of file +* Qin Jin (Renmin University of China, China) + +## Contact +- ASR Challenge - xuankaic@andrew.cmu.edu +- TTS Challenge - jiatongs@andrew.cmu.edu +- SVS Challenge - jiatongs@andrew.cmu.edu +- Research Paper - shinjiw@ieee.org \ No newline at end of file From ce170c65fe5e85f5ff84592f296d18b85baf0424 Mon Sep 17 00:00:00 2001 From: Xuankai Chang Date: Sun, 21 Jan 2024 15:03:08 -0500 Subject: [PATCH 6/6] update asr submission details --- ...rspeech2024-Discrete-Speech-Unit-Challenge.md | 16 ++++++++++------ 1 file changed, 10 insertions(+), 6 deletions(-) diff --git a/_posts/2024-01-19-Interspeech2024-Discrete-Speech-Unit-Challenge.md b/_posts/2024-01-19-Interspeech2024-Discrete-Speech-Unit-Challenge.md index 652bf39d..d1afa27a 100644 --- a/_posts/2024-01-19-Interspeech2024-Discrete-Speech-Unit-Challenge.md +++ b/_posts/2024-01-19-Interspeech2024-Discrete-Speech-Unit-Challenge.md @@ -17,9 +17,9 @@ Description here In conventional speech processing approaches, models typically take either raw waveforms or high-dimensional features derived from these waveforms as input. For instance, spectral speech features continue to be widely employed, while learning-based deep neural network features have gained prominence in recent years. A promising alternative arises in the form of discrete speech representation, where speech signals within a temporal window can be represented by a single index $d$ within a range $K$: $d \in \{1, 2, \dots, K\}$. Three challenging tasks are proposed for using discrete speech representations. - 1. Automatic speech recognition (ASR): We will evaluate the ASR performance of the proposed systems on the proposed data. - 2. Text-to-speech (TTS): We will evaluate the quality of the generated speech. - 3. Singing voice synthesis (SVS): We will evaluate the quality of the synthesized singing voice. +1. Automatic speech recognition (ASR): We will evaluate the ASR performance of the proposed systems on the proposed data. +2. Text-to-speech (TTS): We will evaluate the quality of the generated speech. +3. Singing voice synthesis (SVS): We will evaluate the quality of the synthesized singing voice. Participation is open to all. Each team can participate in any task. This challenge has preliminarily been accepted as a special session for Interspeech 2024, and participants are strongly encouraged to submit papers to the special session. The focus of the special session is to promote the adoption of discrete speech representations and encourages novel insights. @@ -53,10 +53,14 @@ Participation is open to all. Each team can participate in any task. This challe ### ASR Challenge -* Data -* Framework -* Evaluation metrics +* Data: LibriSpeech_100 + ML-SUPERBB 1h set +* Framework: We recommend to use ESPnet for fair comparison. Feel free to let us know your preferrence. +* Evaluation metrics: Word Error Rates (WERs) on 5 test sets. * Submission + * Submission package details: + 1. The discrete speech units corresponding to the test sets in kaldi format. + 2. The predicted transcription corresponding to the test sets. + 3. A technical report in Interspeech2024 paper format (no length limit) ### TTS Challenge - Acoustic+Vocoder