Morris Alper (@MorrisAlper) header
Morris Alper (@MorrisAlper) profile picture

Assistant professor of computer science at @univmiami. My research is at the intersection of multimodal machine learning and linguistics.

🌎
Joined May 2017
Photo by Morris Alper (@MorrisAlper)Photo by Morris Alper (@MorrisAlper)Photo by Morris Alper (@MorrisAlper)Photo by Morris Alper (@MorrisAlper)Photo by Morris Alper (@MorrisAlper)Photo by Morris Alper (@MorrisAlper)Photo by Morris Alper (@MorrisAlper)Photo by Morris Alper (@MorrisAlper)Photo by Morris Alper (@MorrisAlper)Photo by Morris Alper (@MorrisAlper)
The number of languages in the world just got a lot higher! At least constructed ones. Meet ConlangCrafter - a pipeline for creating novel languages with LLMs. A Japanese-Esperanto creole? An alien cephalopod color-based language? Enter your idea and see a conlang emerge. 🧵👇
9
19
19
147
47,096
More cool stuff being cooked up by @eaglescode and @maxmelichov! Now we’ve got trippy articulatory singing voice synthesis in Hebrew!! Enabled by our Renikud G2P conversion model.
CHIMERA sings in Hebrew now!! 🇮🇱 Hebrew is written without vowels so getting this to work was fun :) turn your sound on, details below 🧵
1
8
343
Morris Alper retweeted
I built a voice mode for Claude Code. Like a phone call: speak, listen, interrupt. No push-to-talk. STT and TTS run locally on your Mac (Parakeet + Phonon). One command: /plugin install handsfree --marketplace thewh1teagle/handsfree @bchern @GradiumAI 👇
6
1
6
271
Now accepted to EACL (Findings) 2027! 🇬🇷 Stable Diffusion can draw "Barack Obama" but put in my name and you'll get an imaginary face. Our interpretability method detects whether text-to-image models have memorized identities with only black-box access and no reference photos.
New preprint 💥 What does text-to-image (T2I) models imagine when they hear your name? Our black-box probe measures if a T2I model memorized someone or fabricated a plausible face — no photos or model weights needed. Introducing NAMESAKES 👇
1
10
1,085
Morris Alper retweeted
#SIGTYP is excited to launch a shared task on Collaborative Dictionary Digitization! You can learn more about it here: sigtyp.github.io/st2027-dict… We welcome both data contributors (if you have a dictionary to digitize) and system developers (if you'd like to develop a system)!
4
5
287
Morris Alper retweeted
[1/6] 🔊 We release CLAP2 and LAION-Audio-XL: 365M audio clips (1.39M hours) with 707M captions, used to train five CLAP2 models from scratch. 🏆CLAP2-L reaches 53.6% on MAEB-30, the best CLAP model on the leaderboard. arxiv.org/abs/2611.11592
5
29
3
148
14,313
New preprint! 🐳🦁🐸 Many animals vocalize to communicate, but when we capture them in the field we get noisy recordings. We create a new, lightweight biodenoising model, CallCleaner, that you can run >90x realtime on a laptop CPU! (1/3)
1
7
2
29
1,530
Our key insight is that foundation source separation models (SAM-Audio) can often isolate animal vocalizations but may confuse output streams. By indicating spans to separate with a heuristic animal VAD, we create audio pseudo-labels to distill into a small student model. (2/3)
1
9
207
Check out our project page for audio samples and comparisons! (3/3) Project page: callcleaner.github.io Code: github.com/morrisalp/CallCle… Model: huggingface.co/malper/CallCl…
7
136
התרגומים האוטומטיים הביאו את כל הטוב שבעולם אלי לפיד
8
1
1
391
11,765
Yes, exactly. The result is that papers that make it through are very defensively written so it can deal with all the possible objections. Possible through a slop auto research loop. These are not nice papers for humans to read!
My biggest struggle with reviews in 2026 is that (AI) reviewers don’t seem to know the difference between limitations and critical issues. An LLM can easily find 100+ limitations in any paper (even the best ones in history), but that doesn’t mean the paper shouldn’t be accepted.
3
1
2
51
6,029
Morris Alper retweeted
NeurIPS 2026 Spotlight! 🌟 Physics simulators like MuJoCo MJX are changing robotics. But for studying the sensorimotor learning of speech, there's no equivalent. We built diffTLM, a differentiable vocal tract simulator, to help close that gap. (1/6)
14
58
10
432
45,436
Morris Alper retweeted
Introducing HalachaBench: we gave AI programs from the makers of ChatGPT, Claude, Gemini and others real semicha exam questions from Israel's Chief Rabbinate, in Hebrew, with no books and no internet. The best scored 80 out of 100. Many scored under 11. 🧵
89
80
76
1,157
202,586
Morris Alper retweeted
𝑱𝙪𝒔𝙩 𝙥𝒖𝙗𝒍𝙞𝒔𝙝𝒆𝙙! The 1st of 6 volumes: 𝐷𝑖𝑐𝑡𝑖𝑜𝑛𝑎𝑟𝑦 𝑜𝑓 𝑡ℎ𝑒 𝐼𝑛𝑠𝑐𝑟𝑖𝑝𝑡𝑖𝑜𝑛𝑠 𝑜𝑓 𝐴𝑛𝑐𝑖𝑒𝑛𝑡 𝑁𝑜𝑟𝑡ℎ 𝐴𝑟𝑎𝑏𝑖𝑎 Open Access at this link: tinyurl.com/5n7ht4nj. Nearly 2000 words, with citations and etymological commentary. More soon!
5
98
12
366
21,386
Very interesting taxonomy of LLM writing tics! I'm not sure how exhaustive this is but it definitely includes some points that rub me the wrong way in LLM slop writing...
Replying to @thomashikaru
We collected matched responses from a range of frontier, near-frontier, and pre-frontier models in chat, agentic coding, and essay-writing settings, ranging from Llama-base to Claude Fable 5 and GPT-5.6 Sol.
1
2
288
🎉 Accepted to #ArabicNLP2026! Led by @BenThereDonDhat. Levantine Arabic (LA) is diverse and under-resourced for speech technology. Our new benchmark SHAMS is designed to spur progress in tech for LA across its regional varieties! Details below. @_ArabicNLP
3
2
1
17
862
Key findings: • Levantine-aware systems beat MSA-trained ones, regardless of size • Even the best ASR model gets ~1 in 4 words wrong (>27% WER) • Pronunciation is hardest: audio-to-phoneme is >20% CER
1
1
96