Speech-to-text: Any apps superior to Apple Dictation?

[theorist9]

Honestly, hats off — the depth you've gone into here is going to help a lot of people who stumble on this thread later. Most "which dictation app" discussions stay at the surface; you've actually mapped the terrain, mic differences and all.

Thanks! And it's a rare opportunity for me to be able to discuss this with someone who actually builds these things!

Have you explored mic differences enough to know the point of diminishing returns if you don't have a noisy background?

I tried the BOYA because it's modestly-priced ($40), and is supposed to be a step above cheaper USB mics b/c it has its own A/D converter.

But there are many price tiers above this, including (sampling from Alexa's recommendations on Amazon USA for "Best cost-no-object USB computer desktop microphone for dictation"). For dictating to modern software like Whisper (as opposed to Dragon), Alexa specifically recommends the RØDE NT1 unless your voice varies significantly in volume when you're dictating (which can happen with longer passages), in which case it recommends the Apogee.

Sennheiser Professional Profile - USB Cardioid Condenser ($120)
RØDE NT-USB+ ($150)
Audio-Technica AT2020USB-X Cardioid Condenser USB Microphone ($170)
RØDE NT1 5th Generation Condenser ($220)
Earthworks Icon USB Streaming Mic ($300)
Phillips SpeechMike ($330; handheld, but could be used with stand)
Apogee Hype Mic - USB Microphone ($400)
SpeechWare TBK3 TableMike ($420)

There are even more expensive microphones, but those are designed for studios and live performances rather than computer use, and thus output an analog rather than a digital signal.

And fair catch on Talk Type being cloud — noted. I get that you want strictly no-cloud; my angle's just different. I'm after flexibility, local or cloud per task, rather than local-only as a principle. Neither's wrong, they're just different priorities.

Actually, I've evolved on this. While I would prefer a local-only app, none of the local-only apps I evaluated have the capabiliy I need, so for now I'm resigned to going cloud-based.

From my EDIT 2: "Overall, I found the cloud-based apps are superior to the locally-installed ones for both speed and capability...the best-performing app for me seems to be Aqua Voice, so that's probably the one I'll be purchasing. . " [Aqua Voice is cloud-based only] [And I didn't decide against Talk Type because it's cloud-based. I decided against it, even though it does have real-time dictate anywhere, because it doesn't have Aqua Voice's performance.]

The polish layer: it's good, but I wouldn't claim parity with frontier cloud models. A local ~7B Mistral won't match GPT/Claude on the subtle stuff, and it's noticeably easier for English than for other languages.

Makes sense.

For demanding tasks, no local model I could run, with any amount of RAM, is going to equal the capabilities of the frontier models. But:

1) I assume—please correct me if I'm wrong—that these cloud-based dictation apps aren't polishing my dictation using full-fat frontier models like Opus or Fable since, for a couple hours of dictation/day, the tokens could cost them a significant fraction of their ≈$10/month subscription charge.

If so, my question is: How big a local LLM do I need to match the polishing performance of these cloud-based apps, and how much RAM would I need to run it? Eg., I've read that I could run GPT-OSS-120B or qwen3-next-80b-a3b-instruct with 96 GB RAM. Would those do the job (no idea myself if these would be the best LLM's for dictation polishing) and, if so, would they run fast enough on an M5 Max?

2) Some of these apps, like Spokenly, will connect whatever local dictation model you choose to your user-supplied local LLM. How hard is it to configure a local LLM for dictation polishing, and to connect it to a dictation app, for someone who's never done this before?

[I'm not generally computer-naive—half of my PhD work was computer modelling—but I'm entirely naive about this stuff. To give you an idea how even the basics elude me: I've read you can use Opus for free to test it out. But every time I go onto Anthropic's website to do this, I can only find the paid options! The problem is that all the tutorials out there start at step 3 instead of step 1—the writers incorrectly assume everyone has the basic knowledge they have, so they skip the starting steps.]

 

[genexx]

@theorist9

Mic diminishing returns — I've used three: a cheap webcam mic, my iPhone, and an NT-USB+ (built-in compressor, genuinely nice). All three delivered usable STT results once dialed in. The RØDE wasn't worse than the iPhone — but the iPhone is more forgiving on distance and gain, and since I use it as my webcam anyway, it's become my daily driver upstairs. The RØDE moved down to the studio where I need a dedicated mic. Honestly, for STT the mic matters less than people think. What matters most is how you speak: clear, well-articulated pronunciation beats any mic upgrade. After that, tuning: VAD sensitivity, pause duration, max sentence length. With Parakeet especially, getting the pause threshold right for your personal speech pattern makes a bigger difference than hardware. That's why I expose all of these as user settings. Plus auto-gain that handles different mics automatically — set once, runs.

And here's what surprised me since we last talked: plugging the iPhone in via USB as Continuity camera/mic, after the optimizations your thread kicked off, it was dramatically faster than Wi-Fi. Near-instant. The mic quality was never the bottleneck — the wireless hop was.

My daily setup now: iPhone as Continuity webcam + mic, LLT translating in real time, subtitles going out via OBS Browser Source directly into WebRTC sessions. One device doing camera, mic, and feeding the whole translation pipeline. That workflow drove a lot of the speed improvements in the current build.

LLM sizing — you're right that cloud apps aren't running full Opus on your dictation. Most use smaller, faster models (Haiku/Sonnet tier, fine-tuned 8B–30B) because latency matters more than raw capability, and the token cost would be brutal otherwise. The gap between a good local 7–8B and what these apps actually use is smaller than people assume.

For matching that polish locally: a quantized 30B–70B comfortably covers dictation polishing (punctuation, capitalization, sentence boundaries, homophone resolution). M5 Max with 96GB — plenty of headroom. The 120B models you mentioned work but are overkill; a well-prompted 30B is faster and nearly as good for structured text cleanup.

Connecting a local LLM — the missing step 1 you're looking for: install Ollama (brew install ollama or download from ollama.com), then `ollama pull mistral` in Terminal, and you have a local LLM running with an OpenAI-compatible API at localhost:11434. That's it. 15 minutes. Most apps that support "custom API" can point there.

Anthropic free tier — you're not imagining things, it's not obvious. Go to claude.ai (not anthropic.com), create a free account, chat with Claude directly. Free tier has usage limits but no payment required. anthropic.com is the API/developer side, that's why you only see paid options.

And — LLT giveaway drops tomorrow on MacRumors. 20 full licenses. I'll drop the link here. After all the optimizations your thread kicked off, I'm genuinely curious how it lands in your tables.