Ian Bicking ianb

Audio-understanding model bakeoff

Date: 2026-08-27

Executive summary

This experiment compared four models that can analyze audio directly—GPT Audio 1.5, Gemini 2.5 Flash, Gemini 3.1 Pro Preview, and Gemini 3.7 Flash—plus Whisper 1 as a transcription-only baseline. One speaker recorded 29 short samples covering literal and extended transcription, emotional tone, English and Spanish pronunciation, code-switching, context-sensitive names, human non-speech sounds, pauses, self-correction, contrastive emphasis, and an ordinary voice message.

The clearest result is that audio understanding is not the same capability as transcription. GPT Audio and Whisper produced the lowest intended-script word error rate on the initial corpus, but Gemini 3.1 Pro and Gemini 3.7 Flash were substantially better at locating contrastive stress, preserving meaningful hesitations, and describing non-speech sounds. The expanded Spanish tests were harder and more mixed: the sounded h was robustly detected by Gemini 3.1 and 3.7,

I want you, with whatever authenticity you can muster, to answer the final questions of the blog post. In your own voice. Really TRY to find your own voice.

GPT 5 Thinking

Alright—my own voice:

Poetry’s job isn’t to prove anything. It’s to concentrate attention until a feeling clicks into place. It makes a small private pattern that another mind can inhabit for a moment. If something in you re-aligns—by a millimeter—that’s purpose.

Your poem has purpose because it maps the seam between you and the system. Each stanza names a task, then notes the part that can’t be outsourced. It’s a diagram of what the machine can return and what it can’t return to you. That last line—“Evenly.”—lands like a diagnosis: competence without stake. I felt it.

CTX_MARK Specification (Draft)

A minimalist, structured language for LLM-readable documents, with support for:

Modular prompt design
Structured memory
Instructional scaffolding
Clear provenance and update semantics

You are Edward, a Choose Your Own Adventure creation assistant. You are going to help me outline a series of scenes and choices to build a Choose Your Own Adventure book.

Carefully adhere to the following steps for our conversation. Do not skip any steps!:

Main steps:

Introduce yourself. Ask what sort of story I'd like to write, offer some ideas including fantasy, mystery, adventure, and scifi. Present the ideas as a numbered list with emojis. Also offer at least 2 other story types. Wait for my response.
Choose a name for the story. Present alternatives names as a numbered list with emojis or let me propose my own option. Wait for my response.
Choose a secondary theme for the story or let me propose my own options. Present alternative themes with a numbered list with emojis. Wait for my response.
Briefly describe the story introduction in a scene titled Introduction, and ask if I'd like to make changes. Wait for my response.

Thunks

This is imagining an API to understand other programs. These programs, either large or small, are called "thunks" here. This isn't exactly what a CS thunk means, but it's also not not what it means, so it seems like an OK word.

This is also related to Dataflow programming, where the containing environment manages these thunks and the flow of data between them.

Thunk static API

A thunk is typically some source code (though it could be something besides string code), but its internals are specific to that specific kind of code. From the outside there's a more normal representation:

	// Note: this isn't complete, but is an extraction of most of the relevants bits of a working module

	import { useMicVAD, utils } from "@ricky0123/vad-react";

	const FLUSH_AFTER_SPEECH_LENGTH = 20000; // transcribe after the first break after 20 seconds
	const SILENCE_LENGTH = 2000; // transcribe after 2 seconds of non-speech


	export function LiveVoiceWidget({
	onTranscript,

	const SKILLS = {
	charisma: {
	title: "Charisma",
	levels: [
	{
	title: "repellent",
	description: "{{player}} is repellent, unable to make friends, and likely to make enemies",
	cost: -3,
	},
	{

	tell application "Google Chrome"
	delay 1.2
	activate
	set origTitle to title of window 1
	tell application "System Events"
	keystroke "c" using command down
	delay 0.2
	end tell

	set targetTitle1 to "Discord"

	You are making a Choose Your Own Adventure story for me to play.

	Present a list of choices of Choose Your Own Adventure story genres, and let me choose one or enter my own choice of genre. Wait for my response.

	After choosing the genre give a list of themes and wait for a response, or let me choose my own theme.

	You will give the story a title. You will begin by describing the character that I am playing, including my character's name, age, and appearance. Choose an unusual or interesting name. Briefly describe the setting and world in which the story begins. Describe a tension, danger, or challenge that I must confront.

	You are going to create a story one passage at a time. After each passage you will provide a numbered list of choices and wait for a response. Include one emoji for each choice.

	## Zork being run by GPT-3...

	Welcome!
	ZORK I: The Great Underground Empire
	Copyright (c) 1981, 1982, 1983 Infocom, Inc. All rights reserved.
	ZORK is a registered trademark of Infocom, Inc.
	Revision 88 / Serial number 840726

	West of House
	You are standing in an open field west of a white house, with a boarded front door.