Tips · 模型选择
The biggest speech model is often the wrong choice
"Just use the largest model" is the most common wasted expense in local transcription. Large models are more accurate, but the returns fall off quickly while time, memory and heat rise linearly. More importantly: for Chinese, dialects and noisy rooms, the gap between model families dwarfs the gap between parameter counts.
Language first, size second
Nearly every published comparison is measured on English. If your material is mainly Chinese, that ranking does not transfer — tone, homophones and the ellipsis of spoken Chinese lean heavily on the training distribution. A model built for Chinese (the Tongyi / FunASR family) frequently beats a larger general model on Chinese material despite a third of the parameters.
Cantonese, Hokkien and heavily accented Mandarin are a different step up. There the question is not accuracy but capability: if the language is not covered, no amount of parameters helps — you get fluent-looking nonsense.
Where size helps, and where it does not
Parameters buy two things: rare words and proper nouns, and robustness in noise. Both are real needs — outdoor interviews, conference rooms, jargon-heavy material justify the time cost.
What size does not buy is just as clear: sensible punctuation and segmentation (that is post-processing), timestamp precision (that is alignment strategy), and accent handling (that is training data). Treating a large model as a universal answer means paying for gains where they do not exist.
| Your material | Optimise for | Why |
|---|---|---|
| Clear Mandarin speech, lectures | Mid-size model + word timestamps | Large models add little on clean audio |
| Interviews, panels (multiple speakers) | Diarization capability | Who spoke is harder than what was said |
| Outdoor, live, noisy | Noise robustness | This is where size genuinely wins |
| Dialects, strong accents | Language/dialect coverage | Without coverage, accuracy is moot |
| Jargon-heavy (medical, legal) | Large model + glossary review | Rare words are the strong suit of size |
| Long files (hours) | Speed and memory | Large models hit a memory wall |
Why VAD changes the quality you can see
VAD decides whether there is speech in a stretch of audio. It sounds like an internal detail and is not: segment badly and the model will invent sentences over silence — the classic hallucinated subtitle.
Good VAD cuts long audio at natural pauses so each piece is short enough to transcribe stably. Bad VAD leaves whole paragraphs of text nobody said, which is worse than a typo: while proofreading you have no easy way to tell whether the line was ever spoken.
MonoSub segments with Silero VAD and records each chunk boundary. That is both an accuracy measure and a debuggability one: when something is wrong you can point at a specific chunk instead of distrusting an entire transcript.
Quantization: memory saved, and the price
The same model usually ships at several quantizations (q8_0, q4_k_m, …). The pattern: q8 loses almost nothing against full precision, q4 starts to show, and more aggressive quantization degrades noticeably in noise.
The practical rule: q8 if memory allows, q4 when it is tight, and go lower only when it will not run at all — rather than starting at the smallest to buy speed. Slow is something you can wait out; wrong is something you have to redo.
How to find your model in fifteen minutes
Do not let someone else’s benchmark decide. Use your own material:
- Pick one representative clip containing your worst case — accent, noise, jargon
- Transcribe it by hand as the ground truth
- Run two or three candidate models and compare line by line
- Record not just the error count but how expensive each error is to fix — one wrong proper noun costs more than three wrong particles
- Note the elapsed time too; your minutes are a cost
Fifteen minutes here saves months of second-guessing. And the conclusion only holds for material like yours — which is exactly why no universal recommendation can make the decision for you.
The verdict
Model choice is a trade-off, not an upgrade. Filter by language and capability, choose among what survives by how your material actually behaves, and use size and quantization last to tune performance. Reverse the order and the usual outcome is the slowest model, not winning on the thing you cared about.
MonoSub
Written by the people building a local subtitle tool. What it does and does not do is all in here.
Prices and features follow each vendor’s site; data was checked on the date noted in each post.
Free to download, Pro only for batch
Processing one file at a time is free. Batch work needs a one-time Pro: $9.99 / ¥48 as a Mac App Store in-app purchase, or ¥36 / $8.60 as a direct-store offline licence.
Related reading
MonoSub vs VideoCaptioner: free and open source, or a free app with a one-time Pro
27 Oct 2026 · 5 min read
ComparisonsMonoSub vs MacWhisper: two one-time-purchase local transcribers, compared
27 Oct 2026 · 5 min read
ComparisonsMonoSub vs cloud transcription: 讯飞听见, 通义听悟 and 飞书妙记
27 Oct 2026 · 5 min read
TutorialsYou have local subtitles. Now get them into the finished video.
24 Oct 2026 · 4 min read