Good VS quality

https://quickvc.github.io/quickvc-demo/

https://github.com/quickvc/QuickVC-VoiceConversion

People report Whisper PEFT + LORA tuning gives quite good results:

https://github.com/huggingface/peft/blob/main/examples/int8_training/peft_bnb_whisper_large_v2_training.ipynb

EfficientSpeech, or ES for short, is an efficient neural text to speech (TTS) model. It generates mel spectrogram at a speed of 104 (mRTF) or 104 secs of speech per sec on an RPi4. Its tiny version has a footprint of just 266k parameters. Generating 6 secs of speech consumes 90 MFLOPS only.

https://github.com/roatienza/efficientspeech

https://roatienza.github.io/efficientspeech-demo/

May 12, 2023: Challenge announcement
May 19, 2023: Leaderboard is online and accepting submissions
June 26, 2023: New Language Track Submission Deadline
July 07, 2023: Paper / Model Submission Deadline
July 10, 2023: Paper Revision Deadline

🌍🗣️SUPERB benchmark is back with ML-SUPERB, its multilingual version! The challenge, as one of the #ASRU2023 challenges, includes 3 tracks:
1️⃣ML-SUPERB: For multilingual SSL
2️⃣New language: To new languages!
3️⃣Research: For research papers

More to see 👉 https://multilingual.superbbenchmark.org

Universal Source Separation with Weakly Labelled Data

abs: https://arxiv.org/abs/2305.07447
paper page: https://huggingface.co/papers/2305.07447
github: https://github.com/bytedance/uss

Some people implement streaming speaker diarization manually

https://github.com/pyannote/pyannote-audio/commit/4a6ea9c825b9447a7d03cb9bd94f5f81d661ca16

others just ask ChatGPT to write it

https://github.com/huseinzol05/malaya-speech/commit/564f50c0d91528126fe3b410f387d1b4ff33d364

ChatGPT version is not that bad

The first Arabic TTS Challenge - QASR TTS 1.0 is on!! Register and build your own Arabic Anchor Voice and contribute to enriching #ArabicAI #ASRU2023Challege
More details: https://arabicspeech.org/qasr-challenge/

https://twitter.com/shammur_absar/status/1658429029483986944

Some nice things from industry, autoscaling with Triton and Kubernetes

https://www.speechmatics.com/company/articles-and-news/autoscaling-with-gpu-transcription-models

Multilingual TTS from ElevenLabs

https://twitter.com/radamar/status/1658540025611685888

https://huggingface.co/spaces/elevenlabs/tts

Recent advances in the AudioLM family: 100x higher speed, better consistency, no quality hit - a new paper from and the AudioLM team.

Give it a listen: https://google-research.github.io/seanet/soundstorm/examples/

Arxiv:
https://arxiv.org/abs/2305.09636

Final VoxCeleb Challenge

https://mm.kaist.ac.kr/datasets/voxceleb/voxsrc/competition2023.html

Timeline
May 20th Development set for verification tracks released.
May 31rd Development set for diarisation tracks released.
June 1st Test set released and evaluation server open.
Early August Deadline for submission of results; invitation to workshop speakers.
August 20th Challenge workshop

3 nice Persian TTS datasets

https://www.kaggle.com/magnoliasis/datasets

Whisper is essentially an audio-conditioned LLM. Can we prompt it to do unseen tasks? Introducing PromptingWhisper!

We use simple prompts to adapt Whisper to unseen tasks zero-shot without any finetuning.

📄 Paper: http://arxiv.org/abs/2305.11095
💻 Code: https://github.com/jasonppy/PromptingWhisper

https://twitter.com/csteinmetz1/status/1659458441197355008

I was complaining that LLMs don't have ears... This paper is a solid attempt to try to make that happen.

abs: https://arxiv.org/abs/2305.10790
Work from Yuan Gong et al. at MIT

Interactive demo available https://github.com/YuanGongND/ltu

More details on Soundstorm

https://twitter.com/danlyth/status/1660608450852691968

SoundStorm does a nice job of alleviating a key shortcoming of AudioLM.

By replacing the somewhat cumbersome and slow dual Transformers required for the acoustic token generation, they use bi-directional parallel decoding, leading to a speed-up of two orders of magnitude.

https://arxiv.org/abs/2305.11834

Pengi: An Audio Language Model for Audio Tasks

Soham Deshmukh, Benjamin Elizalde, Rita Singh, Huaming Wang

In the domain of audio processing, Transfer Learning has facilitated the rise of Self-Supervised Learning and Zero-Shot Learning techniques. These approaches have led to the development of versatile models capable of tackling a wide array of tasks, while delivering state-of-the-art performance. However, current models inherently lack the capacity to produce the requisite language for open-ended tasks, such as Audio Captioning or Audio Question & Answering. We introduce Pengi, a novel Audio Language Model that leverages Transfer Learning by framing all audio tasks as text-generation tasks. It takes as input, an audio recording, and text, and generates free-form text as output. The input audio is represented as a sequence of continuous embeddings by an audio encoder. A text encoder does the same for the corresponding text input. Both sequences are combined as a prefix to prompt a pre-trained frozen language model. The unified architecture of Pengi enables open-ended tasks and close-ended tasks without any additional fine-tuning or task-specific extensions. When evaluated on 22 downstream tasks, our approach yields state-of-the-art performance in several of them. Our results show that connecting language models with audio models is a major step towards general-purpose audio understanding