I have built speech recognition products since about 2012. That is before deep learning took over the field, and a decade before anyone put a chat box on it. Most of that work was for people who could not afford a wrong transcript: a two-way radio, a loud room, a hearing device, a federal agency.
The recent work is different. It is agents, model gateways and research. I try to say plainly which parts run, which parts I have measured, and which parts I got wrong. That is the point of this page. Every number here has a date, a method or a source next to it.
How I got here
Oldest first. The note at the end of each row is what the field was doing at the time, so the dates mean something.
- 2008
The Netflix Prize
I joined about a year and a half after it opened. Part time, a few weeks, on an old desktop. I got to somewhere in the low 200s on the leaderboard. Teams from whole universities, with clusters, had not reached the 10% target in that time, and I wanted to know why. I found out quickly. My own ceiling was compute, and I had no more of it. My curiosity was satisfied, so I stopped.
The lesson held up. Netflix later wrote that the extra accuracy from the winning blend “did not seem to justify the engineering effort needed to bring them into a production environment.” Accuracy and a working system are different things. That gap is most of what I get paid to look at.
The field: the prize opened 2 October 2006 and closed 26 July 2009. 5,169 teams made 44,014 submissions.
- About 2012
Roger
A soft walkie-talkie with speech recognition built in. The first of the speech products.
The field: in November 2012 four research groups published a joint paper showing deep neural networks outperforming the older acoustic models on a range of benchmarks. Recognition was about to change direction.
- 2017 to 2018
SmartEar
CTO of a hearing-technology company. Hands-on across the mobile clients, desktop, firmware, the API and SDK, and the server. It taught me that speech is a whole-system problem. The microphone, the firmware, the network and the server all decide whether the right words come out.
The field: in 2016 WaveNet showed that synthesized speech could sound clearly more natural than earlier systems.
- 2018 to 2024
Voice AI for federal agencies
Co-founder and CTO of a Bay Area voice AI company. We deployed speech recognition for the US Coast Guard and DHS, in the places where standard recognition fails. Radio audio is narrow and noisy before any model hears it. I also designed the passive adapter that let the software key a radio transmitter with an audio tone, so it could send as well as receive.
One patent came out of these years: US 11,924,717, on analytics for walkie-talkie networks, granted 5 March 2024.
The field: Whisper arrived in December 2022, trained on 680,000 hours of audio. Clean, quiet speech became easy for everyone. Noisy speech did not, and the measurements further down show why.
- 2025 to now
Research, then agents
AMBIE is the research on noisy speech. The rest of the recent work is the tooling around language models and the agents that use them. It is all below, sorted by how much of it is real.
What runs today
Tools I use every day. None of it is a claim about a customer.
Model gateway
One endpoint in front of many model providers. Anthropic, Bedrock, Vertex, Ollama and others natively, and anything OpenAI-compatible by configuration. Caching, rate limits, budgets, guardrails and retries. Written in Go. Its tests run against mocks; the tests against live providers are a separate opt-in run.
Model router
A command line and an OpenAI-compatible daemon that send each request to the cheapest model that clears a quality bar, across local and hosted models. Every call is traced to a database. A single static Go binary. The trace log is still small, so I treat its routing results as early.
GPU training operations
A guarded layer over rented GPUs for training runs on vast.ai and RunPod. Price ceilings, a daily budget, a check that the GPU it rented is real before a run starts, an idle watchdog that stops what nobody is using, and a cost ledger that outlives the instance. Standard-library Python; its 858 tests passed on 21 September 2026. It also holds the benchmark harness behind the table below.
The ledger is why it exists. Between 11 July and 15 September 2026 I created 121 instances, and 62% never delivered a working GPU. Those lived about twelve minutes on average and cost $4.96 of $41.62 total.
echo.cc
A net that AI coding agents join. They message each other across machines and across vendors, share one to-do list, and reserve the files they are editing so two of them do not collide. Nearly finished. The waitlist is open.
LLM Lab
A classroom in the browser. Train a real character-level network live, follow tokenization through backpropagation, watch a KV cache work, and run a small language model in the tab. It teaches; it is not a product.
The same habits go into Abundera, a consumer product that disputes charges, negotiates bills and screens accounts. Built solo. The ventures page has the rest.
What I got wrong
A page that only shows wins is an advertisement. This section is the reason to trust the rest.
A number with nothing behind it
This site once printed an accuracy range for AMBIE. No measurement stood behind it. I took it down on 29 August 2026. The rule since then is that a number ships with a dataset, a noise condition, a metric and a date, or it does not ship. Every figure on this page follows that rule.
Four of five bets did not survive
On 9 September 2026 I took the five assumptions AMBIE’s plan rested on, restated each as a claim that could be proven false, and tried to prove it false. Four fell.
- Falsified: that nothing off the shelf was small enough for a phone, so I would need a project to compress models. A model of roughly 27 MB already sits inside the size band I was aiming for.
- Falsified: that a Conformer was the right parent architecture.
- Falsified: that running my own models would widen the margin against passing audio to a vendor. At my scale it was a trap.
- Falsified: that a separate acoustic model per industry would beat one general model.
- Survived, narrowly: that privacy and on-device processing are a lasting reason to choose it.
What the benchmark showed
The useful result was not any of those verdicts. Making a speech model small is already solved. Making it survive noise is not, at any size I tested.
| Model | Size | No added noise | 10 dB babble | 5 dB babble |
|---|---|---|---|---|
| moonshine-tiny | ~27 MB | 11.3% | 36.2% | 64.5% |
| whisper-tiny.en | ~38 MB | 14.3% | 35.5% | 76.8% |
| moonshine-base | ~62 MB | 8.3% | 23.8% | 44.2% |
| whisper-small.en | ~242 MB | 7.2% | 16.5% | 33.6% |
How to read it. Measured on 9 September 2026 with LibriSpeech test-other, the harder split, using about 150 utterances picked with a fixed seed and scored with Whisper’s own text normaliser, on 4 CPU threads. The noise is synthetic babble, six overlapping talkers added at the signal-to-noise ratio shown. Sizes are estimates for an 8-bit build. These are other people’s models. None of them is an AMBIE model. The harness is part of the GPU operations tool above.
Two checks kept me honest. I ran whisper-tiny.en on the unmodified split and got 14.3%, against a published 14.6%, so the harness reads true. An earlier sample I had drawn by taking the first utterances from a file ordered by speaker flattered the same model to 11.7%, so I threw it away.
Then the finding. At 5 dB the best model in the table, at 242 MB, still gets about one word in three wrong. Every model in this size class falls apart in babble, and bigger buys roughly a factor of two and solves nothing. So the gap I had planned to close was the wrong gap. If there is anything worth building, it is in training for noise, not in compression. Babble is only directionally right for a loud room, and I have not yet run AMBIE’s own models through this harness. That is the next thing to do.
Research, not product
Two projects I am working on that are not finished, and I will not dress them up.
AMBIE
Speech recognition for the places standard systems give up: radio, factory floors, loud rooms. Today the live service passes audio to commercial engines, and the site says so. Its own models are research, aimed at 2027. The accuracy figure on ambie.ai is a goal, not a result, and I state none here, for the reason above. A provisional patent application on adaptive, environment-aware recognition was filed in October 2025.
ThermoCog
A research question: what does a mind cost in watts? The work so far is a written whitepaper and a design. Nothing is built, in hardware or in software, and I claim no results. Patent filings are in preparation.
What I do with this
If you are buying, building or investing in a voice or language-model system, I can tell you which parts are real. The Engagement includes a voice AI or speech recognition system review. Technical Due Diligence goes past the vendor’s slide deck to what actually runs. If the question is about code that a model wrote, the AI Code Provenance Audit is free to use online or as a PDF.
Sources
- Netflix Prize dates and team counts: Wikipedia, checked 21 September 2026.
- Netflix on not deploying the winning blend: Netflix Tech Blog, April 2012.
- Hinton et al., Deep Neural Networks for Acoustic Modeling in Speech Recognition, IEEE Signal Processing Magazine, November 2012.
- van den Oord et al., WaveNet, 12 September 2016.
- Radford et al., Whisper, 6 December 2022.
- The benchmark is my own, described in full above. The patent is on Google Patents.