Post-doc @ CMU LTI. Audio and speech researcher.
Ancona, Italy
Joined February 2021
- Tweets188
- Following526
- Followers1K
- Likes652
Samuele Cornell retweeted
Work done with my wonderful co-authors around the world: @shinjiw_at_cmu, @forthshinji, Sayaka Shiota, @SamueleCornell, and Satoru Fukayama
Samuele Cornell retweeted
Introducing Monsoon ASR ⚡️ @voicearena_ai
Speech recognition does not have a model problem anymore. It has a data problem.
The best ASR systems are approaching human-level performance in English. But move into the long tail of the world's languages, especially real, conversational speech and error rates can still be 5-10X higher.
Today, we're releasing Monsoon ASR: a new generation of training data built specifically to close that gap.
50 languages. 100,000+ hours. Dense spontaneous speech.
And one goal: Single-digit WER across the world's languages.🧵
Samuele Cornell retweeted
Who wants more speech training data?
YODAS v3 is now available @huggingface
It’s 1.1M hours - the biggest audio dataset ever. And the first at this scale with stereo audio at 48kHz, with timestamped transcripts and translations. 100+ langs
hf.co/blog/espnet/yodasv3
CC-BY-3.0
Samuele Cornell retweeted
🚀 Join the ICASSP 2026 URGENT Challenge! Advance Universal, Robust & Generalizable Speech Enhancement.
🗣 Track 1: Universal Speech Enhancement
🎧 Track 2: Speech Quality Assessment
🔗 urgent-challenge.github.io/u…
#ICASSP2026 #SpeechEnhancement #AI #AudioProcessing
Samuele Cornell retweeted
🗣️ Tomorrow I will be presenting a 𝗗𝗲𝗺𝗼 𝗼𝗻 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝘃𝗲 𝗦𝗽𝗲𝗲𝗰𝗵 𝗘𝗻𝗵𝗮𝗻𝗰𝗲𝗺𝗲𝗻𝘁.
NeurIPS: Saturday, Dec 14, 4:15 pm at West Meeting Room 114, 115
Hope to see you there!
Also feel free to try the Demo for yourself:
🔗 github.com/sp-uhh/gen-se-dem…
If you are still around in Vancouver for Neurips, tomorrow we will have the URGENT challenge workshop from 1.30 pm.
Come by if you are interested in generalizable speech enhancement (also wind will be up to 70km/h tomorrow and is cozy inside 😉)
Lineup: neurips.cc/virtual/2024/comp…
If you are interested in generalizable speech enhancement & restoration consider joining the second edition of the URGENT challenge which will be featured at Interspeech 2025.
It will start this Friday !
We are thrilled to announce the Interspeech 2025 URGENT Challenge, starting on 11/15!
Join us in building universal speech enhancement models to tackle in-the-wild speech data using large-scale, multilingual data. Details: urgent-challenge.github.io/u…
Samuele Cornell retweeted
Hi all,
We have one month before the deadline! Please prepare the submission to our special issue "Multi-Speaker, Multi-Microphone, and Multi-Modal Distant Speech Recognition" at Computer Speech & Language.
sciencedirect.com/special-is…
We're organizing a special issue at Computer Speech & Language about Multi-Speaker, Multi-Microphone, and Multi-Modal Distant Speech Recognition.
Deadline: December 2, 2024
sciencedirect.com/journal/co…
@chimechallenge
In this paper we show that is possible to create synthetic 2-speakers conversations with TTS and LLMs and fine-tune successfully Whisper for multi-speaker ASR generalizing well to real-world scenarios:
arxiv.org/abs/2408.09215
Examples of such synth data: popcornell.github.io/SynthCo…
More details about the "conversational speech" Parakeet TTS model are available here: jordandarefsky.com/blog/2024…
3. This works because we fine-tune Whisper with LoRA.
Thus the acoustic-level mismatch impact is kinda mitigated. There is still however some mismatch that we need to overcome in order to make the approach more scalable: the performance plateaus after 80h of synth data.
@WavLab
2. We show (in our setting) that using a totally synthetic dataset is the same as using legit real-world data but from another domain.
E.g. using synthetic data for fine-tuning and then test on mixer6 affords same performance as using Fisher data.
Some takeways:
1. If we use generated LLMs transcripts in place of original target domain transcripts (e.g. Fisher) does not impact much the performance. This can change for some domains where there is significant specialized jargon (healthcare) but we can use LLMs to augment
Samuele Cornell retweeted
⭐⭐ Call for task proposals! CHiME 9! @chimechallenge
Come and share your ideas about distant microphone speech processing with us!
Task proposals deadline October 25.
chimechallenge.org/current/c…
Samuele Cornell retweeted
We release a detailed paper, model weights (model and codec) and streaming inference for Moshi! Beyond the model itself, we believe our findings will be useful to audio language models. "Inner Monologue" for the win!
Today, we release several Moshi artifacts: a long technical report with all the details behind our model, weights for Moshi and its Mimi codec, along with streaming inference code in Pytorch, Rust and MLX. More details below 🧵 ⬇️
Paper: kyutai.org/Moshi.pdf
Repo: github.com/kyutai-labs/moshi
HuggingFace: huggingface.co/kmhf
Samuele Cornell retweeted
I'm excited to announce @WavLab's XEUS - an SSL speech encoder that covers over 4000+ languages!
XEUS is trained on over 1 million hours of speech. It outperforms both MMS 1B and w2v-BERT v2 2.0 on many tasks.
We're releasing the code, checkpoints, and our 4000+ lang. data! 🧵
If you are interested in generalizable speech enhancement that can tackle "speech-in-the-wild" data, different sampling rates and is able to restore audio from different distortions check this out.
We have a new challenge at NeurIPS 2024.
Website: urgent-challenge.github.io/u…
We are thrilled to announce the URGENT 2024 Challenge - a new speech enhancement (SE) competition at NeurIPS 2024: urgent-challenge.github.io/u…
This challenge aims to unify diverse distortions and sampling frequencies using a single universal SE model. #URGENT2024 (1/4)
Samuele Cornell retweeted
Hi all,
This is the third call for papers about the SynData4GenAI workshop.
Good news!
While the submission data was originally due on June 18th, we'll extend it to June 24th.
Please submit your papers at syndata4genai.org
We look forward to your submissions!
This is a very cool work on speech separation in real world settings (AMI meeting).
It is robust because it is weakly supervised (diarization labels only) and thus can be trained on real world data, unlike most neural methods which have to rely on synthetic data.
Our paper entitled "Neural Blind Source Separation and Diarization for Distant Speech Recognition" is accepted to Interspeech 2024!
Our neural FCASA is a method to jointly separate and diarize speech mixtures without supervision by isolated signals.
arxiv.org/abs/2406.08396