An interactive installation that puts you through a machine version of LADO — the real-world system governments use to determine asylum seekers' origins by listening to their accents. Built with Whisper, Streamlit, and a lot of uncomfortable questions.
This started as a course reading material Michelle Pfeifer's chapter "The Native Ear: Accented Testimonial Desire and Asylum" from the book Thinking with an Accent (2023) about machine listening. I was going to build something about speech recognition bias — interesting but safe.
Pfeifer describes a system called LADO — Language Analysis for the Determination of Origin. Since the 1990s, European governments have been using linguistic analysis to verify asylum seekers' claimed origins. The process works like this: a person flees their country, arrives at a border, and claims asylum. They might not have a passport. They might not have any documents at all. So the state listens to them speak and decides, based on accent alone, whether they're really from where they say they're from.
The premise is brutally simple: how you speak = where you're from. And therefore: where you're from = whether you deserve protection.
The problem, as Pfeifer and many linguists have pointed out, is that this premise is wrong. People migrate. People grow up in multiple countries. A Somali woman raised in Kenya might speak with Kenyan pronunciation. LADO could — and in a real court case Pfeifer documents, did — deny her asylum because she "sounded Kenyan." The system confuses accent with origin, and origin with identity.
What hit me hardest was the concept Pfeifer calls "the native ear" — the idea that accent isn't located in the speaker but in the listener. It's not about how you talk. It's about how someone else hears you talk, and what they assume from that hearing. The listener's biases, not the speaker's identity, determine the outcome.
I wanted to build something that lets people feel this — not read about it, but experience being judged by a machine ear.
The installation is structured as a three-step test:
Step 1: The Biased Listener. The system randomly selects an audio clip in a foreign language and generates "misheard lyrics" — the foreign speech re-written as phonetic nonsense in your native language. If you set your native language to Chinese and get a Japanese audio clip, you'll see something like "抠哇 填其嘎 依卡拉，多抠卡 嘿 得卡开 太呆死 内" — which is Japanese re-heard through Chinese phonetics. This is the "空耳" (soramimi) concept: the same sound, filtered through a different linguistic ear, becomes something entirely different.
Step 2: Human Accommodation. You record yourself reading these misheard lyrics aloud. This is where it gets interesting. You're reading text in your native language's characters, but the text was designed to phonetically approximate a foreign language. How do you read it? Purely in your native pronunciation? Do you unconsciously drift toward the original language's sounds? Do you deliberately try to imitate the foreign pronunciation?
Step 3: Forensic Judgment. The machine (OpenAI's Whisper model) analyzes your recording. It runs language detection — what language does it think you're speaking? — and transcribes what it hears. Then it compares its transcription to the original true text of the foreign audio. The result is a "Machine Conviction Score": a composite of how strongly the machine identified a language in your voice and how closely your speech matched the original. Zero means the machine didn't even recognize the target language. One hundred means it's fully convinced you are a speaker of that language.
The final output is a "linguistic passport" — a pie chart of detected language probabilities, a conviction score, and a machine verdict that ranges from "passport denied" to "passport granted."
Between recording and analysis, there's a two-question self-assessment: