Parroting Journal · Articles

Mac Voice Typing Gets Names and Jargon Wrong? Fix It in 3 Layers

After a while with voice typing, the most annoying problem usually isn't a sentence it can't understand. It's the same few words going wrong every time: a coworker's name, a product name, a piece of jargon. This article splits “making voice typing understand you” into three layers. It explains what each layer can and can't fix, and shares the numbers from training a model on my own recordings.

· About 9 min read

繁體中文English简体中文日本語
A parrot beside a Mac, fixing misheard words one by one

I'm Sona Orbit, the independent developer who builds Parroting on my own, so this is about my own product. The training numbers in this article come from my own voice. That is one person's data, so your results may differ. Feature descriptions match Parroting 0.4.13.

1. The short version: pick the layer that matches the error

LayerBest forWhere you do itHow fast it worksCan't fix
Layer 1
Personal dictionary
Fixed names, product names, English abbreviations, e.g. Jmini → GeminiOn your Mac. Confirm after you edit the textStarts working on the next sentenceHomophones that depend on context (e.g. there / their); sounds that were never picked up
Layer 2
Personalize tone hints
Hints that are too sensitive, e.g. getting flagged when you tease a friendOn your Mac, after 20 pieces of feedbackTrains in seconds, adopted only if it passes validationOnly changes whether a hint shows. Doesn't change the words it recognizes
Layer 3
Custom speech model
Jargon you say often, accents, mixed-language speech, wrong homophonesExport your data, train on a PC with an NVIDIA GPU, import it back to your MacNeeds recordings and training timeGets slightly worse at understanding other people

Most people are fine with the first two layers. Consider the third only if you want to push the error rate down further.

2. Why the same words keep going wrong

Speech recognition works in two steps. First it turns sound into “what this sounds like.” Then it picks the most common spelling among all the words that sound the same. So errors are rarely random. They show up in a few predictable places:

  • Uncommon names: “Kathryn” is rarer than “Catherine,” so the model keeps picking “Catherine.”
  • Product names and jargon: A word the model has never seen gets spelled as the closest sound it knows. Gemini becomes Jmini.
  • Homophones: “there” and “their” sound identical. Only the surrounding words can tell them apart.
  • Accent and run-together speech: When you talk fast and words blur together, the first step already goes wrong.

With the first two, the model heard you correctly and just picked the wrong spelling. A dictionary fixes that. With the last two, the model itself needs to get used to your voice, and a dictionary can't help.

3. Layer 1: Personal dictionary

The most direct fix is to write down the words that keep going wrong. After each transcription, the wrong spelling gets swapped for the right one. Parroting's dictionary runs on your Mac and works even with AI turned off. If you use AI cleanup, Parroting also passes these words to it.

Auto-learning dictionaries have a common problem. You rewrite a whole sentence in different words, the software assumes it misheard you, and it stuffs unrelated words into the dictionary. It gets messier over time. Parroting handles it this way:

  • Sound comes first: When you paste and then edit the text, Parroting compares how the words sounded before and after. Pairs that sound alike (for example “Catherine” → “Kathryn”) go to the confirm list on the “Personalization” page. Pairs that sound very different go under “Reworded.” They never enter the dictionary and are never used for training. If it sorted one wrong, press “Actually misheard” or “Not misheard, I reworded it” to fix it.
  • Only what you confirm counts: Press “Add to dictionary” in the list, and it takes effect from the next sentence.
  • You can add words yourself: Settings → Words & history → Dictionary → “Add.” Write the correct spelling and the common mistakes. You can also import a CSV (first column is the correct spelling, second column is the mistake).
A parrot swaps a card with a misspelled name for the correct one and files it in the dictionary

What a dictionary can't fix: A dictionary means “when you see A, replace it with B.” Words like “there / their,” where both spellings are right depending on context, can't be swapped directly. And if you talk so fast that the first step never caught the sound, the dictionary has nothing to replace. Those go to layer 3.

4. Layer 2: Tone hints, tuned on your Mac

Speaking carries emotion more easily than typing. Parroting includes a small English tone model of about 14 MB (Chinese has its own, about 11 MB). It runs on your Mac. It reads the sentence you just said and the context (which app you're in, what you were just talking about) and judges seven things at once: scene, audience, purpose, formal or not, tone, whether to show a hint, and whether it sounds too casual. No API key, no internet.

But everyone draws the line in a different place. If you say “You idiot, late again lol” to a friend of ten years, that's normal banter. If a hint pops up every time, you'll want to turn it off. So this layer adjusts to your feedback:

  1. If a hint is wrong, press “Hint was wrong” or “Mute hints in this app.”
  2. Training unlocks at 20 pieces of feedback: Open the “Personalization” page in the sidebar, review the feedback, and press train under “Personalize tone hints.” It finishes in a few seconds on this Mac. Your original model stays untouched.
  3. It can't be adopted unless it passes validation: Out of every 5 pieces of feedback, 1 is held back from training and used only for checking. Built-in test sentences also confirm that sentences which should trigger a hint still do. If it fails, Parroting tells you this training would break sentences that were previously right.
  4. You can undo it anytime: After you press “Adopt,” the version history is kept. Press “Revert” to go back to the previous version or to the original model.
A parrot holding a small balance scale, adjusting how sensitive hints should be

Keep in mind that this layer only affects tone hints. It doesn't make the recognized words any more accurate.

5. Before layer 3: Run a personal accuracy check

Before you think about training a model, find out where Parroting mishears you. Settings → Personalization → “Read aloud” is the personal accuracy check. One round is 10 sentences and takes about 2 minutes. You read from a script. After each sentence, Parroting marks the words it misheard, and at the end of the round you get your error rate.

If a marked word sounds the same as the word in the script, Parroting picked the wrong spelling and you didn't misspeak. Errors like that can usually be improved with the dictionary or a custom model. The answers for these recordings are known to be correct (they're the script you read). They stay on your Mac until you delete them, and you can use them for training later.

6. Layer 3: Custom speech model, tested

Parroting's default speech recognition model is SenseVoice Small. I used the open-source training kit parroting-asr-training to fine-tune it with LoRA, which trains only about 2% of the parameters. I tested it on my own dictation recordings. Every test clip was a recording the model never saw during training.

Training kit results (2 hours of everyday dictation, 263 test clips)

ModelError rate
Original model6.7%
Training run 14.7%
Training run 25.5%

Errors dropped by 20 to 30 percent. The biggest gains were on names, product names, and mixed-in English. But two runs on the same recordings differed by almost a full percentage point. The reason is that the “answers” for both runs were transcribed by a model, so they contain errors of their own, and results bounce around. That's why you should always test on recordings the model hasn't seen before deciding whether to use it.

Earlier experiments: more recordings, very little gain

In two earlier experiments, the answers were chosen by comparing several models and letting an AI judge pick. They were not checked by a human either:

ModelTraining recordingsError rate
Original modeln/aAbout 5.0%
Version 1About 30 minutesAbout 3.6%
Version 2About 2 hoursAbout 3.4%

With 4 times the recordings, the error rate dropped only another 0.2 percentage points. My conclusion: correct answers matter more than a large pile of recordings. Read-aloud recordings and corrections you've confirmed have correct answers, so they make the best training material.

A parrot wearing headphones listens to recordings while a chart beside it shows errors falling

The cost, and the pitfalls I hit

  • It gets worse at other people: On a general test set (FLEURS), version 1's error rate rose by only about 0.2 percentage points. Version 2's rose by about 1.2. The more familiar the model gets with you, the worse it does on everyone else. It's your own Mac, so the trade is usually worth it, but know the cost.
  • English words typed twice: Version 1 had a side effect where English words repeated. Google became Googlegle and email became emailmail. Version 2 had far fewer. After training, don't just check the error rate. Use the model for a few days and watch for new quirks.
  • Answers must be Simplified Chinese: This matters if you dictate Chinese. I tried training directly on Traditional Chinese answers, and the model got worse the more it trained. SenseVoice's vocabulary is mostly Simplified Chinese, so the kit converts answers to Simplified automatically. When Parroting outputs text, it converts back to the characters you use (for example, Traditional Chinese with Taiwan usage).

7. Training your own: process and cautions

Parroting itself doesn't train speech models. You train on your own computer with an NVIDIA GPU. Your voice still never gets uploaded to me.

  1. Prepare recordings: You need at least 20 to 30 minutes of recordings with correct text. Read-aloud recordings are the most reliable. Corrections you've confirmed in the confirm list count too, but the recording has to still be saved to export it. By default Parroting keeps only the latest 300 recordings. To keep more, go to Settings → Advanced → History → “Keep recordings” and set a longer time.
  2. Export: Settings → Words & history → Data location → My data → “Export training data…” (it's also on the Read aloud card).
  3. Train: Use parroting-asr-training (MIT license, documentation in English). It needs Linux, or Windows with WSL2. An 8 GB GPU is enough, and an RTX 4090 takes about 8 to 9 minutes. CPU-only works too but is slow: for 90 minutes of recordings, a 16-core machine takes about 50 minutes per epoch. After training, the kit compares before and after on the test clips, and the report says “Use it” or “Don't use it” outright.
  4. Import back into Parroting: Settings → Talk → Speech model & language → “Manage models” → “Import model…”. Parroting first has the Mac's built-in voice read a test sentence aloud, and only adds the new model to the list if it can transcribe it.
  5. Compare before and after: At the top of the “Manage models” window, use “Test with your read-aloud recordings.” It uses only the 1/5 of your read-aloud recordings held back as the test set, so sentences the model trained on can't inflate the score. Test once before importing and once after. If you don't like the result, press “Back to the standard model.”

8. FAQ

Do the personal dictionary and tone hint personalization need a GPU or internet?

No. Both layers run on your Apple silicon Mac. You don't need internet, an API key, or a GPU. Only layer 3, the custom speech model, needs to be trained on a computer with an NVIDIA GPU (or a very patient CPU).

Are my recordings uploaded?

No. Parroting's recordings, speech recognition, dictionary, and tone personalization all happen on your Mac. A file leaves Parroting only when you press “Export training data…” yourself, and where it goes is up to you.

Will Parroting get more accurate over time even if I don't train a custom model?

Yes. A dictionary entry takes effect on the next sentence once you confirm it. Fixed errors like names and product names are the most common and the easiest to fix. Try a 2-minute read-aloud round first to see your error rate and where the errors are, then decide whether to go on to layer 3.

How many recordings do I need to train a custom model?

At least 20 to 30 minutes of recordings with correct text. My tests show that correct answers matter more than the amount of audio. With 4 times the recordings, the error rate dropped only another 0.2 percentage points.

What if the training makes things worse?

For the speech model, press “Back to the standard model” in “Manage models” to return to the original. For tone hint personalization, you can press “Revert” at any time to go back to the previous version or the original model.

← Back to articles