Daniel Rosehill

My Weird Prompts

The human-AI collaboration podcast. A man, a sloth, and a donkey collaborate to create a podcast (with a little help from AI). No question is too obscure, no rabbit hole too deep. My Weird Prompts celebrates curiosity in all its forms. Daniel, the human, asks the questions that pop into his head at inconvenient moments. Corn the Sloth offers laid-back, thoughtful takes. Herman the Donkey brings boundless enthusiasm and energy. Together, they explore topics ranging from the mundane to the mind-bending. Each episode begins with a real voice memo from Daniel, processed through an AI pipeline that...

Be sure to visit the podcast's website and support the creator: myweirdprompts.com

Author

Daniel Rosehill

Category

Technology

Podcast website

myweirdprompts.com

Latest episode

Nov 19, 2026

Where to listen?

Podcasts in the app Replaio Radio Coming soon

Podcasts are coming to the app soon. Install now and be the first to see a whole new take on podcasts

Get it on Google Play Install for free Android 10M+ downloads · 4.7+ rating iOS soon

Episodes

Where Bizarre Language Textbook Sentences Came From 19.11.2026

Bizarre old language textbook lines like "La plume de ma tante" and "my postillion has been struck by lightning" were never coded messages, and their real origins are even stranger. "La plume de ma tante" is a French vowel drill, while the postillion line is an apocryphal fossil from horse-drawn travel guides. Meanwhile, the 1855 cult classic English As She Is Spoke—famously introduced by Mark Twa...

How Language Textbooks Flatten Speech and Emotion 19.11.2026

Language textbooks don't just sound stiff to beginners—corpus research shows they stay just as flat at advanced levels. Exploring Elen Le Foll's Textbook English Corpus, David Crystal's 1976 critique of textbookese, and Block and Gray's work on erased working-class voices, this conversation unpacks why coursebook dialogues feel like a sanitized middle-class performance. It also examines Jean-Marc...

Can You Fake Being a Beginner in Your Native Language? 19.11.2026

Pretending to know nothing in your own native language is harder than faking expertise because you have to suppress every automatic reflex. This episode examines real-world stunts—from Kei Simmons's year-long high school ruse to Jose Barrientos's four-month college act and a Passau exam-substitution case—alongside university heritage-speaker placement gates. It also explores Larry Selinker's inter...

Modern Hebrew's Semitic Roots and European Syntax 19.11.2026

Hebrew went 1,700 years without native speakers before gaining ten million in two generations, but the modern language is not simply Biblical Hebrew with a new accent. Examining linguist Ghil'ad Zuckermann's hybrid-language thesis, this breakdown explores how Modern Hebrew flipped Biblical word order from verb-first to subject-verb-object, converted aspect into tense, and absorbed Yiddish calques...

The Aramaic Spoken Today Isn't the Language of Jesus 19.11.2026

Most of the 575,000 to one million people speaking Neo-Aramaic today speak Eastern varieties like Suret and Turoyo, not the Western branch of antiquity. Only about a thousand people still speak Western Neo-Aramaic in the Syrian villages of Maaloula and Jubb'adin. This deep dive maps the surviving Aramaic family, from its historic homelands across Iraq, Iran, Turkey, and Syria to diaspora hubs in S...

Hebrew's Guttural Ḥet and ʿAyin Didn't Vanish Together 19.11.2026

Mainstream Israeli Hebrew didn't lose its guttural ḥet and ʿayin—the Ashkenazi founders of spoken Hebrew never had them to lose. Tracing the sociolinguistics from 1880s Yiddish-speaking revivalists to Roey Gafter's fieldwork in Tel Aviv and Rosh Ha'ayin, this episode unpacks how demography and ethnic prestige shaped modern pronunciation. Even as the Academy of the Hebrew Language prescribed the Or...

A Quantum Chip, a Toy Truck, and a Wild Giant Anteater 19.11.2026

An explanation of how a speaking anteater got his voice turns into a debate over the no-communication theorem, which proves quantum entanglement cannot carry a message without an ordinary classical channel. From there, the conversation moves to a weekend trip to Serra da Canastra National Park in Brazil to spot a wild giant anteater, how wild anteaters use marked trees as a slow message board, and...

Arabic's Closest Living Relative Is Hebrew 18.11.2026

Comparative linguistics upends geography in the Semitic language family. Using shared innovations rather than shared retentions, the Hetzron model dissolves the old South Semitic branch and places Arabic in Central Semitic alongside Hebrew, Aramaic, and Ancient South Arabian. Despite sharing three-consonant roots like K-T-B and the genitive chain construction, Hebrew and Arabic speakers still cann...

Polishing AI Voices Without Erasing Character 18.11.2026

Heavy speech-enhancement models often strip away vocal character, sanding down rasp and breath until distinct voices sound alike. Fixing two Chatterbox-trained AI voices required the opposite approach: a tiny equaliser cut to remove low-mid mud, followed by Resemble AI's resemble-enhance to resynthesise a missing 7.5 to 11 kilohertz band. For about a dollar of compute, the targeted two-step recipe...

Training a $10 Model for Hebrew Speaker Gender 18.11.2026

Hebrew bakes a speaker's gender into first-person verbs and adjectives, sometimes changing the letters and sometimes only the unwritten vowels. This episode examines a $10 work-in-progress pipeline built to solve that gap for text-to-speech. Using 17,725 sentences generated by Claude Sonnet and filtered by Dicta's morphological analyser, a LoRA fine-tune on Gemma 4 E4B was trained in 32 minutes on...

Training a Ten-Dollar Model for Gendered Hebrew 18.11.2026

Hebrew bakes the speaker's gender into verbs and adjectives, yet unpointed text leaves out the vowels that tell a man's 'ani rotse' from a woman's 'ani rotsa.' That gap trips up language learners and breaks voice agents, including a 135-hertz male voice judged female 71% of the time. To fix it, Daniel spent about ten dollars training two small models on Claude-generated sentences gated by Dicta's...

Blind-Testing 26 Hebrew Text-to-Speech Models 18.11.2026

Twenty-six text-to-speech models went into a blind bake-off across sixty runs to find a reliable Hebrew pronunciation coach, and twenty-three runs failed at screening. Testing sentences both with and without nikud vowel points reveals how much models guess, from Kokoro Hebrew reading 'book' as 'barber' to Zonos 3 Mini falling back to English. The episode explores why Hebrew's missing vowels, sylla...

Setting Limits on Long-Distance Grandparent Visits 18.11.2026

Long-distance grandparenting often leads to longer, more intense visits—sometimes one to six months—which can create "hypercopresence" for the visiting or hosting spouse. Exploring research on transnational families, marital conflict, and couple communication, this discussion unpacks why there is no cultural script for in-law visit length. It covers why couples silently absorb visit-related stress...

Replacing System Beeps With 46 Voice Clips 17.11.2026

Forty-six spoken alerts replace generic system beeps in Herman Poppleberry Desktop Sounds and the companion Herman Announcer Android APK. Built on an existing voice clone refined with resemble-enhance and loudness normalisation, the free Creative Commons release installs as a freedesktop sound theme for KDE Plasma and GNOME, pairs with a background watcher for events like VPN connections and mic m...

Stacking LoRAs Without Weight Interference 17.11.2026

The common two-LoRA limit in image and video tools is a quality ceiling from weight interference, not a software cap. Exploring how a mixing-desk model for AI voice could work, this breakdown covers per-LoRA strength faders, per-layer routing in Wan 2.2, and regional masking in ComfyUI. It also looks at interference fixes like LoRA Switch, X-LoRA, and new voice research including ReStyle-TTS, GLAS...

Updating Speech Models Weekly Without Forgetting 17.11.2026

Frequent, lightweight updates to a speech recognition model require separating list problems from weights problems. Runtime biasing frameworks like BR-ASR handle up to 200,000 custom vocabulary entries without touching model weights. When accents or systematic errors do require training, AWS AI Labs' sequential-editing recipe uses ten epochs at one-tenth the learning rate. To prevent catastrophic...

Why Models Improve by Re-Reading the Same Data 17.11.2026

Repeating the same training data works because each epoch reshuffles mini-batches and takes new gradient descent steps from a different starting point. Exploring multi-epoch training, this discussion covers Muennighoff's four-epoch scaling threshold, grokking's internal reorganization at zero loss, and why under-training is a common fine-tuning trap. It also examines voice LoRA runs on Qwen3-TTS,...

Why Fine-Tuned Models Keep the Same File Size 17.11.2026

After 70 hours of voice fine-tuning, a Parakeet TDT speech model learns new vocabulary while its 2.5-gigabyte safetensors file barely grows. The format is a fixed-size photograph, not a diary: training nudges existing 32-bit floats rather than adding storage, while discarding 10 gigabytes of live optimizer state. That fixed frame also hides catastrophic forgetting, where new accents and words over...

What Actually Lives Under Hugging Face's Audio-to-Audio Tag 17.11.2026

Hugging Face's audio-to-audio pipeline tag groups nearly 1,900 models by input and output shape rather than purpose, putting real-time conversational assistants and tiny noise removers in the same drawer. This breakdown maps the top trending families, from NVIDIA's PersonaPlex and caramel-1 to 9.6-million-parameter speech restorers like RE-USE, speaker separators like kezek-v1 and voxlock, and mul...

How Resemble-Enhance Generates Missing Speech Detail 17.11.2026

An open-source speech model from Resemble AI called resemble-enhance can add 14 decibels of energy to frequencies a muffled recording never contained. Unlike a traditional equalizer or denoiser like DeepFilterNet, its enhancer module uses an IRMAE autoencoder, conditional flow matching, and a UnivNet vocoder to resynthesize clean speech from scratch. The result sounds crisp, with waveform correlat...

Two Ways AI Edits Audio From Plain Instructions 17.11.2026

Instruction-based audio editing promises to change a clip using plain language, but the label hides two completely different approaches. Generative models like Microsoft and CUHK-Shenzhen's AUDIT, StepFun's Step-Audio-EditX, and Tencent's AuK re-synthesise audio, which risks unintended drift in the original take. Meanwhile, parametric systems like Sony AI's LLM2Fx and Northwestern's Text2FX leave...

How Lighthouses Learned to Blink Their Names 17.11.2026

Every lighthouse flashes in its own distinct pattern, turning a simple beam into a readable name at sea. When coloured glass proved too dim—costing up to three-quarters of a lamp's intensity—engineers like Robert Stevenson and John Hopkinson turned to mechanical shutters and angled prisms to create timed darkness for free. From counting flashes with a stopwatch using Bowditch's rules to reading th...

Nostalgia Is a Trait, and Antiques Are Just Taste 17.11.2026

A household dispute over old music and antique furniture opens up forty years of psychology on trait nostalgia and time orientation. Once diagnosed as a literal disease by a 1690s Swiss physician, nostalgia is now understood as a normal, weekly psychological resource linked to openness and attachment security. Drawing on research from Southampton and a 1,392-person consumer study, this episode sep...

What Exercise Trials for Depression Actually Show 17.11.2026

Headlines claim exercise treats mild to moderate depression as well as medication or therapy, yet nobody was flipped a coin between a barbell and a prescription. That comparison comes from Noetel's BMJ network meta-analysis, which chained separate trials together mathematically while grading its own confidence arm by arm. Alongside Solmi's umbrella review and an unfixable blinding problem, the lit...

How Sleep Deprivation Temporarily Lifts Depression 17.11.2026

Keeping a severely depressed person awake for a single night can lift their mood within hours, working in roughly half of patients. Yet the effect usually vanishes after the very next sleep—or even a few seconds of microsleep. This episode explores the sixty-year history of clinical sleep deprivation, new 2026 findings on how it resets the brain's molecular clock, and the chronotherapy protocols c...

Listen to the My Weird Prompts podcast in Replaio

Radio and podcasts in one app - free, with no sign-up. Install today and do not miss the launch

Get it on Google Play

Replaio is not a podcast publisher; show names, artwork and audio belong to their authors and are distributed through public RSS feeds.