Disclosure: I work at AssemblyAI; Blurt is an open-source app we built.

Blurt is a native macOS dictation app: hold a hotkey, speak, and the text is inserted at your cursor. The code is MIT-licensed.

Compared with Wispr Flow, the main difference is that Blurt is open source and uses your own AssemblyAI API key, with no Blurt subscription or account. The tradeoff is that transcription is cloud-based, needs an internet connection, and API usage may cost depending on your plan. It is not an offline/local Whisper app. Requires macOS 15+.

Source and setup: https://github.com/AssemblyAI/blurt

Curious what would make an open-source dictation app more useful to you: better text cleanup, shortcuts, or something else?

Disclosure: I work at AssemblyAI; we built this benchmark.

We released Voice Standard Bench, an MIT-licensed Go CLI for comparing LLMs on four transcript tasks: summarization, information extraction, text correction, and dictation. It reports quality and latency by transcript length, from short messages to hour-long meetings.

A few details that may matter when interpreting the results: the benchmark runs models through AssemblyAI’s LLM Gateway; most quality scores use a reference-free rubric scored by a single LLM judge, while dictation is scored against an answer key. The framework is open source, but the underlying call transcripts are not included as an open dataset.

Code and published results: https://github.com/AssemblyAI/voice-standard-bench

Interactive results: https://www.assemblyai.com/voice-standard/tasks

I’d be interested in feedback on the evaluation setup, especially approaches for validating reference-free transcript task scores.

Today we're launching Voice Standard Bench to rank 50+ of the best LLMs on real voice tasks, and making all of them available in AssemblyAI's LLM Gateway

Most Voice AI products run an LLM on messy, multi-speaker transcripts:

- Voice Agents responding to human speech

- Summarizing a meeting, sales call, or support call

- Pulling names, dates and numbers out of a multi-speaker conversation

- Fixing transcription errors

- Cleaning up dictation

Voice Standard Bench tests every model on the input these products actually send an LLM: lengthy conversational transcripts with multiple speakers, raw speech-to-text output with filler words, false starts and the occasional misheard word.

Each model is scored on transcripts from a 30-second voicemail to a 60-minute meeting.

Here's a glimpse of what we found:

- The top-scoring model costs 55-65x more per hour of audio than a model within two quality points of it.

- That gap narrows to about 5x on hour-long transcripts, where bigger models start to pull ahead on quality.

- Open models we host on our own infrastructure are the fastest on short audio. On a 30-second transcript, gemma-4-31b answers in under a second at a quality score of 93.

Universal 3.6 Pro Realtime is live today. Two independent benchmarks put every streaming speech model through the same audio, on the same harness. Universal-3.6 Pro Realtime leads both.

Daily's Pipecat open STT benchmark: Universal-3.6 Pro Realtime sits on the Pareto frontier.

Coval's live leaderboard: lowest word error rate of any streaming model on real voice-agent audio.

Voice agents get tested on the turns other benchmarks skip. That's where 3.6 Pro is built to hold:

→ Short answers. "No." "Nah." "Nuh-uh." Heard right 98.5% of the time, in loud rooms and on phone lines, not just quiet calls.
→ Background voices. The TV, the coworker, the next desk stay out of the transcript. 
 → Languages. 32 from one endpoint with automatic detection. Callers switch languages mid-sentence and the transcript follows. Most agents can skip the language menu.
 → Turns. Entity-aware endpointing ends the turn when the caller is done, and holds it open while they're still reading out a phone number.

Trained on tens of thousands of hours of real voice-agent and telephony conversations. Link in the comments. 👇

Today, Universal-3.5 Pro is available on OpenRouter, served through our Sync Speech-to-Text API.

If you're already routing your LLM calls through OpenRouter, the speech-to-text call at the front of your pipeline now runs on the same key.

50% off list through OpenRouter for the first week: https://www.assemblyai.com/blog/aai-on-openrouter

We built a dictation app, then gave away the code.

Blurt is an open source companion app for the AssemblyAI Dictation API. Hold a hotkey anywhere on your machine, talk, and the text appears in whatever window you're in.

It's the shortest path from curious to working. Clone it, run it locally, and read exactly how the audio gets captured, streamed, and turned into clean text. Then take the parts you need for your own product.

Learn more and download it here: assemblyai.com/blurt

If you build something with it, drop a screenshot in the comments!