15:I[981,["658","static/chunks/658-a698693f56d0116a.js","183","static/chunks/183-6beeeb801c274aaa.js","474","static/chunks/app/%5Btype%5D/layout-631c4730b83c8223.js"],"Posts"] 16:T734,Curious to hear what approaches people are taking, what the bottlenecks are, and whether anyone here is pushing toward the goal of "AI that understands you, the first time."

I've been diving into the gap between benchmark ASR performance and real-world speech. Models like Whisper and Deepgram show impressive >95% accuracy in ideal conditions. But in the wild — accents, noisy environments, emotional speech, code-switching, overlapping speakers — accuracy often drops sharply, often to the mid-80s or worse.

This matters because the next wave of AI won't be chatbots; it will be hands-free, real-time systems in contexts like:

- care work (voice logs) - crisis communication - home healthcare - security rounds - field operations - "I need help" micro-interactions

In these high-stakes contexts, 85% accuracy means critical information gets lost.

What seems missing today:

- Fine-tuning pipelines for noisy, accented speech - Reinforcement learning loops (user corrections → model improvements) - Fast per-speaker adaptation - Better handling of disfluencies ("uh," "um," repairs) - Scaling-law insights applied to ASR models - Evaluation metrics that reflect real environments instead of curated datasets

What I'm trying to understand:

- What prevents ASR from reaching reliable >99% accuracy in real-world conditions? - Is the bottleneck the model architecture, data quality, or something else?

Would love to hear from anyone who has:

- Worked on Whisper fine-tuning - Tackled multilingual or accented ASR - Shipped speech systems in noisy environments - Developed conversational (not dictation) ASR models - Built correction-feedback training loops - Deployed ASR in safety-critical or field environments

What worked? What failed? What surprised you?

Why Fears of a Trillion-Dollar AI Bubble Are Growing

Austria's armed forces rely on LibreOffice instead of Microsoft

Cultural Death in the Age of the Excessive Choice

Pg_stat_insights: PostgreSQL Performance Monitoring Extension

The missing link in JavaScript tools

Pro-tax Austrian heiress gives away $27M of her inheritance

Losing Confidence

Show HN: MTXT – Music Text Format

The 30-foot sea cow quickly hunted to extinction because of its tasty meat

Celebrity Birthdays: Toddler Edition 2025 (Z-Image)

Exploiting open Ollama instances for free LLM inference

Shrinking While Linking

Silicon Valley's Man in the White House Is Benefiting Himself and His Friends

Why is real-world ASR still ~85% when lab models claim >95%?

Open source YouTube Content Blocker based on words

Elf a modern CLI helper for Advent of Code written in Python

UserScanner

Israel's IDF Bans Android Phones: iPhones Now Mandatory

Great_Consciousness_Theory_2025

The Space of Minds

Show HN: Let Claude Code call other LLMs when it runs in circles

Triton Plugins

The NaK Population: a 2019 Status [pdf]

In-Depth Analysis of the Latest Deep Research Technology

Self-hosting my photos with Immich

Show HN: YOLO11 and OpenCV Object Detection in Java – Full End-to-End Tutorial

Reviewing live-bootstrap and reflections on trusting trust

How to detect text which has been written by ChatGPT

AI Can Generate Code. Is That a Threat to Computer Science Education?

Show HN: I cross-compiled llama.cpp to run on Windows XP

Why Fears of a Trillion-Dollar AI Bubble Are Growing

Austria's armed forces rely on LibreOffice instead of Microsoft

Cultural Death in the Age of the Excessive Choice

Pg_stat_insights: PostgreSQL Performance Monitoring Extension

The missing link in JavaScript tools

Pro-tax Austrian heiress gives away $27M of her inheritance

Losing Confidence

Show HN: MTXT – Music Text Format

The 30-foot sea cow quickly hunted to extinction because of its tasty meat

Celebrity Birthdays: Toddler Edition 2025 (Z-Image)

Exploiting open Ollama instances for free LLM inference

Shrinking While Linking

Silicon Valley's Man in the White House Is Benefiting Himself and His Friends

Why is real-world ASR still ~85% when lab models claim >95%?

Open source YouTube Content Blocker based on words

Elf a modern CLI helper for Advent of Code written in Python

UserScanner

Israel's IDF Bans Android Phones: iPhones Now Mandatory

Great_Consciousness_Theory_2025

The Space of Minds

Show HN: Let Claude Code call other LLMs when it runs in circles

Triton Plugins

The NaK Population: a 2019 Status [pdf]

In-Depth Analysis of the Latest Deep Research Technology

Self-hosting my photos with Immich

Show HN: YOLO11 and OpenCV Object Detection in Java – Full End-to-End Tutorial

Reviewing live-bootstrap and reflections on trusting trust

How to detect text which has been written by ChatGPT

AI Can Generate Code. Is That a Threat to Computer Science Education?

Show HN: I cross-compiled llama.cpp to run on Windows XP

UserScanner