Pulled 811 AssemblyAI blog posts, their full Hacker News history (77 submissions), 12 weeks of cross-industry trend data (550 companies), and Deepgram's HN history as a category control. Here's what the data actually says before the ideas.
Recent launches (Jun–Aug 2026): Universal-3.5 Pro Realtime (Jun 23, first streaming STT that takes the agent's question as input), Contextual Awareness (Jul 2), Universal-3.5 Pro async with native code-switching + joint diarization (Jul 7), the Sync API (Jul 15, transcript in a single HTTP call at ~134ms), Coval "Human Parity Zone" result at 3.40% WER (Jul 25), Summarization + Action Items rebuilt on the speech_understanding object (Aug 12), and Agent Context Carryover for LiveKit (Aug 13). Plus Medical Mode, LLM Gateway, and the $4.50/hr Voice Agent API earlier in the year.
...
Trend data: MCP is the runaway #1 topic (1,248 normalized mentions, 93 companies), AI Agents #2 (1,060), Real-time #4 (925). Voice AI sits at 890 mentions across only 24 companies — the narrowest field of any major topic, meaning low competition per unit of attention. Multi-agent systems is one of only 7 topics trending up (+59.7% WoW). Local AI spiked +500% in late July and is being covered by just 6 companies. Emerging terms with absurd velocity: "scorer" (+7,500%), "watermarking" (+5,200%).
Here are the five, each pairing a proven viral format with a recent launch.
-
"How Speaker Diarization Actually Works"
A from-scratch, math-and-diagrams explainer of the full pipeline: VAD → segmentation → speaker embeddings → clustering → the overlap problem → why joint transcription-and-diarization models changed the architecture. No product until a single line at the bottom. Illustrated with audio you can play alongside the embedding space.
Why it works: This is a carbon copy of the format that produced their four biggest HN hits ever — "How DALL-E 2 Works" (252 points), "How Imagen Works" (142), "Introduction to Diffusion Models" (98), "What is residual vector quantization?" (45). Diarization is genuinely under-explained on the open web, and Universal-3.5 Pro's joint diarization gives them first-hand authority. Their recent diarization posts (Aug 13, three of them) are all comparison/SEO framed and will not travel; the underlying expertise is there and being wasted.
-
"DER Ranks Diarization Systems Backwards. Here's the Proof."
Take the 248-word teaser from July 21 and turn it into the full argument: why the industry-standard Diarization Error Rate is time-weighted and therefore systematically forgives the errors humans actually notice (short back-channels, speaker merges), with an open-source repo that re-ranks published systems under cpWER and shows the ordering flip. Include the audio clips where DER says "good" and your ears say "wrong."
Why it works: Deepgram's two highest-signal non-explainer posts were exactly this genre — "Lies, Damn Lies, and Benchmarks" and "OpenAI's Text Cleaning Hides Whisper's True Word Error Rate." Contrarian methodology critiques with reproducible code are HN's single most reliable format because they invite verification rather than trust. AssemblyAI already has the cpWER Python implementation shipped (Aug 13) and a benchmark position to defend. Risk to manage: it must re-rank including cases where AssemblyAI loses, or it reads as marketing and dies.
-
"We Built a Streaming ASR That Runs Entirely in Your Browser Tab"
A build-your-own weekend project: quantized encoder, chunked audio, WebGPU inference, a hand-rolled turn detector, no server, no API key. Ship the repo and a live demo. End with an honest table of what breaks — accents, entity errors, latency variance — and where a hosted model is the right call.
Why it works: "Build Your Own Imagen Text-to-Image Model" hit 111 and MinImagen was submitted separately; the build-it-yourself format is proven for this specific audience. Local AI spiked +500% in the trend data (week of Jul 20) and is being covered by only 6 of 550 tracked companies — an unusually open lane. LogRocket published a browser voice-agent piece on Aug 6, which confirms pull rather than saturation. The honest limitations section is what makes the Sync API pitch land instead of feeling like a bait-and-switch.
-
"Anatomy of a One-Second Voice Turn"
A millisecond-by-millisecond teardown of everything between a user's last syllable and the agent's first audible word: mic buffer, network, VAD/endpointing, STT finalization, LLM TTFT, TTS first chunk, playback. Ship an open measurement harness so readers can profile their own stack and post their numbers.
Why it works: Real-time is the #4 trend overall (925 mentions, 104 companies) and one of the few holding steady while everything else declines. This directly monetizes the Sync API launch (134ms single-call transcripts) and the Coval 3.40% WER / fastest-response-time result without either being the subject. Performance-budget teardowns with a shareable tool are highly reproducible on social — people post their own numbers, which is distribution you don't have to buy. It also reframes the "latency vs accuracy" debate they've already written about twice into something concrete.
-
"Give Your Coding Agent Ears: An MCP Server for Audio"
Build and open-source an MCP server that lets Claude Code or any agent transcribe, diarize, and query audio as a tool — then use it for something non-obvious, like having an agent debug its own voice-agent failures by listening to the call recordings.
Why it works: MCP is the #1 trend across all 550 tracked companies by a wide margin (1,248 mentions, 93 companies — roughly 4x the company breadth of Voice AI), and multi-agent systems is one of only seven topics actually trending up. AssemblyAI's 2025 MCP explainer was HN-submitted, so the audience overlap is proven. They already have the raw material: "Prompting Claude to build voice agents" (Jun 24), the LLM Gateway call-scoring post, and the voice-agent-regression diagnostic pipeline (Jul 18). Show HN with a working repo is the highest-conversion launch surface available, and it's the one format where being a product company is an asset rather than a liability.
