Monday, 16 November 2026
20 episodes · 6596 in all
#6638: How Python Wheels Work Without Running Code
A Python wheel is just a ZIP archive with a .whl extension, and installing one runs zero code—it only copies files.
#6637: EQ, Compression, and Limiting for Spoken Voice
Aim for three to six decibels of gain reduction on your loudest words, and always cut muddy frequencies before you compress.
#6636: The Five-Step EQ and Compression Chain for Any Voice
A clean voice chain takes just five moves—high-pass, mud cut, presence, de-ess, and compression—applied in strict order before the limiter.
#6635: Why Consistency Beats a Perfect Studio Setup
A desk reflection arrives 0.6 milliseconds late, carving an 855 Hz notch right through the middle of the vocal range.
#6634: Ancient Israelite Men Went Bareheaded
For most of Jewish history, there was no rule requiring men to cover their heads—constant covering emerged in Sasanian Babylonia, not Sinai.
#6633: Voice Engineering and the TTS Recording Matrix
Low-rank adaptation can boost zero-shot speaker similarity by up to 30 percentage points, even when trained on noisy or synthetic audio.
#6632: Building a Toddler-Proof Linux Laptop With One App
A decade-old Debian package called BamBam turns any old laptop into a toddler-proof toy by running as its own isolated login session.
#6631: Why a $30 Cardputer Can't Replace a Pocket Linux PC
The $30 M5Stack Cardputer ADV packs a keyboard and microphone into a wallet-sized shell, but it can't run a browser or act as a USB mic.
#6630: Using LLMs as Annotators for Silver Training Data
Using a seed set to test 66 prompt configurations beats stuffing fixed few-shot examples into an LLM annotator.
#6629: How Canopy Searches AI Skills and Tools on Demand
Loading 927 tool names cost 20,010 tokens before a single prompt, so Canopy puts skills and tools behind one search endpoint.
#6628: Same 100 Hours, 98% vs 6.74% Word Error in ASR
The same 100 hours of LibriSpeech audio yields a 98.4% word error rate in SoralASR-23M-Preview and 6.74% in k2-fsa's pruned stateless transducer.
#6627: Separating Voice From Pronunciation in a TTS Model
A forty-clip x-vector test proved that a text-to-speech model stores Hebrew pronunciation in its weights, completely separate from speaker voice.
#6626: How a Wooden Mouse Proved Machine Learning Worked
In 1950, Claude Shannon built a wooden mouse that learned a 25-square maze using telephone relays and solved it in 18 seconds.
#6625: How MP3 Deletes 90% of a Song Without You Noticing
MP3 shrinks CD audio by up to 95 percent by exploiting auditory masking, a quirk of human hearing first observed in 1894.
#6624: How One Matrix Trick Powers Text, Image, and Voice LoRAs
A LoRA adapter shrinks a 350-gigabyte GPT-3 checkpoint down to 35 megabytes by adding two tiny matrices to the frozen weights.
#6623: Teaching Five AI Voices Israeli Hebrew
Span-masking 461 code-switched takes taught five synthetic voices Israeli Hebrew for about $4.40, tested in a ten-minute radio comedy.
#6622: How Far Postgres Actually Scales in Production
ChatGPT's backend serves 800 million users from a single unsharded Postgres primary, using nearly fifty read replicas and strict timeouts.
#6621: How the Kaldi Team Pivoted to OmniVoice
The k2-fsa team at Xiaomi went from a two-language TTS proof of concept to OmniVoice's 600-language model in just ten months.
#6620: When Speech Enhancement Erases Podcast Voice Character
Cleaning up podcast voices with NVIDIA's RE-USE model before training LoRA adapters accidentally erased the hosts' rough edges and personalities.
#6619: When Merging LoRA Adapters Stops Being Exact
Folding a half-mebibyte LoRA adapter into a base model still produces a 32-mebibyte dense matrix—and bf16 rounding can silently erase the update.