ParrotingPersonalization
The more you use it, the more it gets you
Parroting learns three things about you: words it mishears, where your lines are, and your voice.
Dictionary
Fix a misheard word once; it is right next time.
Works from the next sentence💬Personalize tone hints
Hints that match your sense of tone.
Trains after 20 items🎙️Read aloud and your own model
Read a few sentences and see where Parroting mishears you; the recordings can train your own model.
Checkup works now · training needs an NVIDIA GPULayer 1
Misheard once, right next time
When you fix a misheard word, Parroting suggests adding it to the dictionary. It is added only when you confirm.
Details
- Parroting tells mishearings from rewordings by how the words sound. Mishearings can go into the dictionary; rewordings get their own tab and never go into the dictionary or training. You can correct its call.
- Confirmed items fold away under "Confirmed (N)", so the list shows only what is left.
- You can also add words yourself: Settings → Words & history → Dictionary → "Add", with the right spelling and common mishearings.
- Best for names, product names and technical terms.
- It cannot fix sound-alike words that depend on context; your own speech model helps there.
Layer 2
Hints that match your sense of tone
If a hint is off, press "Hint was wrong". After 20 items, Parroting trains a layer of judgment that is only yours, on your Mac.
Details
- Confirm your feedback under "Personalize tone hints" on the "Personalization" page in the sidebar. Nothing is uploaded.
- It only affects tone hints, never speech recognition. Each language you give feedback in gets its own section.
- One in five items is held back to check the result, plus a built-in check set; if the new layer does worse, it is not used.
- Every adopted version is kept, so you can go back one version or to the original model.
- Works in English and Chinese.
This sounds a bit harsh; take another look before sending
Layer 3
Read a few sentences, see where it mishears you
"Read aloud" is your personal recognition checkup: rounds of 10 sentences, about 2 minutes. Each misheard word is marked as soon as you finish a sentence, and you get your error rate at the end of the round. These recordings have exact text, so they can also train a speech model that is only yours.
Advanced: train your own speech model on your recordings. Two runs by the developer with the kit (2 hours of everyday dictation, tested on 263 clips it never saw):
19–30% fewer errors, mostly on names, product names and mixed-in English you say often. Both runs used the model's own transcripts as text, which have mistakes; in an earlier test, going from 30 minutes to 2 hours of audio only moved the error rate from 3.6% to 3.4%. So read-aloud clips with exact text matter more than hours of audio. The trade-off: slightly worse on other people's speech (about 1 point). Training runs on your own computer with an NVIDIA GPU; Parroting itself never trains a speech model.
- Read aloud + confirmed fixesSettings → Personalization → "Read aloud", a few rounds; corrections you confirmed in Review count too
- Export training dataSettings → Words & history → Where your data is → My data → "Export training data…", also on the Read aloud card
- TrainWith the open-source parroting-asr-training, about 8–9 min on an RTX 4090
- ImportSettings → Talk → Speech model & language → "Manage models" → "Import model…"
- Test with your read-aloud recordingsAt the top of "Manage models": "Test with your read-aloud recordings", before and after importing, and compare the error rates
For people who train: full steps and notes
- Exact text matters most: read-aloud recordings are the most reliable, then corrections you confirmed. Recognition-only text works too, but hits a ceiling.
- "Export training data…" writes a zip (parroting-asr-dataset v1). Read aloud and confirmed corrections are selected by default; "Recognition only" is optional. A fixed 1 in 5 is kept as a test set and never trained on. Recordings over 30 s are skipped.
- Read-aloud recordings are kept until you delete them. Confirmed corrections need their recordings: Parroting keeps the latest 300 by default; to keep more, raise Settings → Advanced → History → "Keep recordings".
- The training kit needs Linux or Windows with WSL2, works on 8 GB GPUs, is documented in English and is MIT-licensed. Have at least 20–30 minutes of recordings with correct text. It fine-tunes SenseVoice Small with LoRA, about 2% of the parameters. CPU-only works but is slow: about 50 minutes per epoch for 90 minutes of audio on 16 cores.
- "Test with your read-aloud recordings" only uses the 1-in-5 test set, so sentences you trained on do not inflate the result.
- The kit converts Chinese text to Simplified for training: Traditional text made it worse in testing. Parroting converts the output back to your script.
- Steps:
ptrain.py checkchecks the data,ptrain.py traintrains and compares before and after on the test set (the report says "Use it" or "Don't use it"),export_gguf.shwrites a .gguf of about 250 MB. - On import, Parroting tries one sentence and enables the model only if it recognizes words; "Back to the standard model" any time.
- License: SenseVoice Small is under the FunASR model license, and so is anything trained from it.
- Privacy: the export holds your voice and what you said; think before sharing it.
- Training on a Mac has not been verified yet; use a computer with an NVIDIA GPU.
Ready to make Parroting yours?
Download the free Mac app and let it get better the more you speak.
Download for Mac ↓Free · Apple silicon · macOS 11 or later · v0.4.13 · ~50 MB