Zhihang Xie
@zhihangxie.bsky.social
🚀 New paper: Speech Discrete Tokens or Continuous Features?
📄 aclanthology.org/2025.emnlp-m...
🧩 A comprehensive benchmark of SpeechLLMs using HuBERT/WavLM with Qwen & LLaMA.
✨ Continuous features outperform overall, while discrete tokens excel at phoneme-level detail.
📄 aclanthology.org/2025.emnlp-m...
🧩 A comprehensive benchmark of SpeechLLMs using HuBERT/WavLM with Qwen & LLaMA.
✨ Continuous features outperform overall, while discrete tokens excel at phoneme-level detail.
aclanthology.org
November 12, 2025 at 2:37 PM
🚀 New paper: Speech Discrete Tokens or Continuous Features?
📄 aclanthology.org/2025.emnlp-m...
🧩 A comprehensive benchmark of SpeechLLMs using HuBERT/WavLM with Qwen & LLaMA.
✨ Continuous features outperform overall, while discrete tokens excel at phoneme-level detail.
📄 aclanthology.org/2025.emnlp-m...
🧩 A comprehensive benchmark of SpeechLLMs using HuBERT/WavLM with Qwen & LLaMA.
✨ Continuous features outperform overall, while discrete tokens excel at phoneme-level detail.
Reposted by Zhihang Xie
🚀 Exciting news from the @fbk-mt.bsky.social group!
@bsavoldi.bsky.social , @linaconti.bsky.social, @matteo-negri.bsky.social & @luisabentivogli.bsky.social are attending #EMNLP2025 in Suzhou 🇨🇳!
Come to our sessions & let's connect:
🔗 mt.fbk.eu/fbk-mt-at-em...
We’re also hiring postdocs!⚡
@bsavoldi.bsky.social , @linaconti.bsky.social, @matteo-negri.bsky.social & @luisabentivogli.bsky.social are attending #EMNLP2025 in Suzhou 🇨🇳!
Come to our sessions & let's connect:
🔗 mt.fbk.eu/fbk-mt-at-em...
We’re also hiring postdocs!⚡
November 4, 2025 at 9:02 AM
🚀 Exciting news from the @fbk-mt.bsky.social group!
@bsavoldi.bsky.social , @linaconti.bsky.social, @matteo-negri.bsky.social & @luisabentivogli.bsky.social are attending #EMNLP2025 in Suzhou 🇨🇳!
Come to our sessions & let's connect:
🔗 mt.fbk.eu/fbk-mt-at-em...
We’re also hiring postdocs!⚡
@bsavoldi.bsky.social , @linaconti.bsky.social, @matteo-negri.bsky.social & @luisabentivogli.bsky.social are attending #EMNLP2025 in Suzhou 🇨🇳!
Come to our sessions & let's connect:
🔗 mt.fbk.eu/fbk-mt-at-em...
We’re also hiring postdocs!⚡
🚀 SimulMEGA: MoE Routers as advanced policy makers for Simultaneous Speech Translation 🎧🌍
Mixture-of-Experts routing → smarter decisions on when & how to translate, balancing latency vs quality in real-time speech. Paper link at arxiv.org/pdf/2509.012...
Mixture-of-Experts routing → smarter decisions on when & how to translate, balancing latency vs quality in real-time speech. Paper link at arxiv.org/pdf/2509.012...
arxiv.org
September 3, 2025 at 7:33 AM
🚀 SimulMEGA: MoE Routers as advanced policy makers for Simultaneous Speech Translation 🎧🌍
Mixture-of-Experts routing → smarter decisions on when & how to translate, balancing latency vs quality in real-time speech. Paper link at arxiv.org/pdf/2509.012...
Mixture-of-Experts routing → smarter decisions on when & how to translate, balancing latency vs quality in real-time speech. Paper link at arxiv.org/pdf/2509.012...
🚀 AdvST: Adversarial training aligns speech and text distributions without parallel data! Combines adversarial learning + hidden-state swapping to fix length mismatch & boost low-resource speech translation. ieeexplore.ieee.org/document/108...
Adversarial Speech-Text Pre-Training for Speech Translation
Large-scale pre-training has been shown to benefit speech translation tasks. However, existing multimodal pre-training efforts rely on parallel corpora for semantic alignment, potentially limiting per...
ieeexplore.ieee.org
July 9, 2025 at 10:05 AM
🚀 AdvST: Adversarial training aligns speech and text distributions without parallel data! Combines adversarial learning + hidden-state swapping to fix length mismatch & boost low-resource speech translation. ieeexplore.ieee.org/document/108...
🚀 Boost rare-phrase translation in speech! Uses **bilingual dictionaries** (e.g., "climate change"→"Klimawandel") to dynamically bias outputs.
✅ **+21%** recall in streaming ST
✅ **+85%** in multimodal LLMs
🔗: arxiv.org/abs/2506.09175
✅ **+21%** recall in streaming ST
✅ **+85%** in multimodal LLMs
🔗: arxiv.org/abs/2506.09175
PHRASED: Phrase Dictionary Biasing for Speech Translation
Phrases are essential to understand the core concepts in conversations. However, due to their rare occurrence in training data, correct translation of phrases is challenging in speech translation task...
arxiv.org
July 2, 2025 at 8:42 AM
🚀 Boost rare-phrase translation in speech! Uses **bilingual dictionaries** (e.g., "climate change"→"Klimawandel") to dynamically bias outputs.
✅ **+21%** recall in streaming ST
✅ **+85%** in multimodal LLMs
🔗: arxiv.org/abs/2506.09175
✅ **+21%** recall in streaming ST
✅ **+85%** in multimodal LLMs
🔗: arxiv.org/abs/2506.09175
Reposted by Zhihang Xie
🔍 Stiamo studiando come l'AI viene usata in Italia e per farlo abbiamo costruito un sondaggio!
👉 bit.ly/sondaggio_ai...
(è anonimo, richiede ~10 minuti, e se partecipi o lo fai girare ci aiuti un sacco🙏)
Ci interessa anche raggiungere persone che non si occupano e non sono esperte di AI!
👉 bit.ly/sondaggio_ai...
(è anonimo, richiede ~10 minuti, e se partecipi o lo fai girare ci aiuti un sacco🙏)
Ci interessa anche raggiungere persone che non si occupano e non sono esperte di AI!
Qualtrics Survey | Qualtrics Experience Management
The most powerful, simple and trusted way to gather experience data. Start your journey to experience management and try a free account today.
bit.ly
June 3, 2025 at 10:24 AM
🔍 Stiamo studiando come l'AI viene usata in Italia e per farlo abbiamo costruito un sondaggio!
👉 bit.ly/sondaggio_ai...
(è anonimo, richiede ~10 minuti, e se partecipi o lo fai girare ci aiuti un sacco🙏)
Ci interessa anche raggiungere persone che non si occupano e non sono esperte di AI!
👉 bit.ly/sondaggio_ai...
(è anonimo, richiede ~10 minuti, e se partecipi o lo fai girare ci aiuti un sacco🙏)
Ci interessa anche raggiungere persone che non si occupano e non sono esperte di AI!
Reposted by Zhihang Xie
📢 Come and join our group!
We offer a fully funded 3-year PhD position:
📔 Automatic translation with large multimodal models: iecs.unitn.it/education/ad...
📍Full details for application: iecs.unitn.it/education/ad...
📅 Deadline May 12, 2025
#NLProc #FBK
We offer a fully funded 3-year PhD position:
📔 Automatic translation with large multimodal models: iecs.unitn.it/education/ad...
📍Full details for application: iecs.unitn.it/education/ad...
📅 Deadline May 12, 2025
#NLProc #FBK
Reserved topic scholarships | Doctoral Program - Information Engineering and Computer Science
iecs.unitn.it
April 22, 2025 at 10:13 AM
📢 Come and join our group!
We offer a fully funded 3-year PhD position:
📔 Automatic translation with large multimodal models: iecs.unitn.it/education/ad...
📍Full details for application: iecs.unitn.it/education/ad...
📅 Deadline May 12, 2025
#NLProc #FBK
We offer a fully funded 3-year PhD position:
📔 Automatic translation with large multimodal models: iecs.unitn.it/education/ad...
📍Full details for application: iecs.unitn.it/education/ad...
📅 Deadline May 12, 2025
#NLProc #FBK
ReShape Attention bridges speech & text models without extra parameters. Achieves +8.5% BLEU in translation by leveraging acoustic cues, outperforming cascade/E2E methods. Efficient & scalable. Check the paper by Kano et al. (2025) at: ieeexplore.ieee.org/stamp/stamp.....
IEEE Xplore Full-Text PDF:
ieeexplore.ieee.org
April 9, 2025 at 3:04 PM
ReShape Attention bridges speech & text models without extra parameters. Achieves +8.5% BLEU in translation by leveraging acoustic cues, outperforming cascade/E2E methods. Efficient & scalable. Check the paper by Kano et al. (2025) at: ieeexplore.ieee.org/stamp/stamp.....
New research fuels the debate between cascaded and E2E speech translation! The challenge of error propagation is addressed by incorporating multiple ASR candidates, along with HuBERT features to preserve acoustic information lost after ASR. Check the paper by Min et al. at: arxiv.org/pdf/2502.00377.
arxiv.org
February 6, 2025 at 10:18 AM
New research fuels the debate between cascaded and E2E speech translation! The challenge of error propagation is addressed by incorporating multiple ASR candidates, along with HuBERT features to preserve acoustic information lost after ASR. Check the paper by Min et al. at: arxiv.org/pdf/2502.00377.