跳到正文
DEV Community· Utterwove·· 36 分鐘前AI 評分45

Android 離線聽寫:解讀裝置端 Whisper

Offline Dictation on Android: On-Device Whisper Explained

AI 導讀

裝置端 Whisper 可在 Android 手機上處理語音轉文字,模型下載到手機後,聽寫毋須連接伺服器;模型較大通常辨識較準,但會佔用更多儲存空間和記憶體,速度亦較慢。Utterwove 提供可下載的多語言 Whisper Tiny、Base 等模型,個人字典則可用詞語替換修正人名及專有名詞。下載模型、登入、試用和授權檢查仍需網絡,選用的 AI 寫作模式亦會把文字傳送給服務提供者。

正文

TL;DR: "Offline dictation" means the speech is turned into text on your phone, not on a server. Whisper is an open speech model that makes this possible. Smaller models are faster and lighter. Larger ones are usually more accurate but heavier. Names and jargon are where any model slips, and a personal dictionary is the cheap fix. Some things, like signing in or optional AI rewriting, still use the internet.

For developers: A plain-language look at running Whisper-style speech models on a phone, and where the trade-offs are.

What "offline" actually means

Most voice typing works like this: your phone records audio, sends it to a server, the server returns text. That is fast and often accurate, but it needs a connection, and your audio travels.

On-device dictation moves that step onto the phone. A speech model is downloaded once. After that, the phone does the listening and the converting by itself. You can switch on airplane mode and it still works, as long as the model and your language are supported.

I make a dictation app (Utterwove), so I will say plainly what this does and does not give you. It gives you independence from signal and a simpler privacy story for the dictation step. It does not make every feature offline. More on that below.

Whisper in plain words

Whisper is a speech recognition model that OpenAI released as open source in 2022. It was trained on about 680,000 hours of audio in many languages, so one model can handle speech from many languages and accents. Because the weights are public, developers can run it directly on laptops and phones. Projects such as whisper.cpp make that practical.

Whisper comes in sizes. The names run from tiny and base up to small, medium and large. Bigger means more parameters, which generally means better recognition, a larger file, more memory and slower results.

Model size What it is like Good for
Tiny Smallest and fastest; makes more mistakes Quick notes, older or lower-memory phones
Base A step up in accuracy, still light A sensible default to try first
Larger models Usually more accurate; slower, more storage and memory Longer or harder dictation, if your phone copes

Utterwove's Android app offers downloadable on-device models, including multilingual Whisper Tiny and Base and others such as Parakeet. The exact list can change, so check the model screen in the app. Language support varies by model and device.

Two other terms come up. English-only models are trained for English and are often a little better at it. Multilingual models cover many languages, including Hindi and Telugu, but they are not equally strong in all of them.

Setting it up in practice

  1. Download a model on Wi-Fi. Model files can run from tens of megabytes to well over a hundred. Check your storage first.
  2. Pick the language you will speak, then the matching model.
  3. Test in airplane mode with a sentence you know. If it works there, it will work on a train.
  4. Compare two models with the same short paragraph. Do this three times. One good result proves little.

Why names go wrong, and the fix

A speech model predicts the most likely words. It has seen "Priya" less often than "pre-ya", and it has never seen your client's company name. So the same name comes out wrong every time. Larger models help a little, but they do not know your world.

A personal dictionary solves this in a plain way. You tell the app: when you hear this phrase, write that. In Utterwove on Android, the dictionary is a list of phrase replacements. It does not retrain the model. Examples I would add on day one:

  • Your own name and your manager's name
  • Your company and product names
  • Acronyms spelled the way you want them
  • Your email sign-off or a booking link

It takes a minute to add a name, and then you stop correcting the same word. There is a longer guide on dictionary names on our site.

What still needs the internet

Be wary of any app that says "100% offline" without qualification. In ours:

  • Downloading a model needs a connection once.
  • Sign-in, the trial and the licence check use the network.
  • Optional AI writing modes send text to a provider you choose, and that provider may charge separately. Local Email and Journal formatting does not need one.

According to Utterwove's Google Play Data safety section, no data is shared with third parties. You can also delete your account at any time from the privacy page.

The honest trade-offs

On-device is not always better. A big cloud model can beat a small local one on some voices and languages. Local models use battery and storage. A tiny model on a cheap phone can be fast and wrong. I do not claim better accuracy than any other app, and I would be suspicious of anyone who did without a published test. The practical answer is to try your own voice, in your own noise, with your own words.

FAQ

Does Whisper need internet?
Not once the model file is on your phone. You need a connection to download it, and some app features, such as sign-in, still use the network.

Is a bigger model always better?
Usually it is more accurate, but it is also slower and uses more memory. On a modest phone, a smaller model that responds quickly may feel better to use.

Will offline dictation understand Hindi or Telugu?
The multilingual models list both. Quality varies with the model, device, voice and background noise, and Telugu and Hindi generally have less training data than English. Test with your own sentences.

Is offline dictation more private?
For the dictation step, your audio does not have to leave the phone, which is a simpler setup. Other features such as sign-in or provider-based AI rewriting may still use the network, so read each app's data policy.

Try it

If you want to test on-device dictation on Android, Utterwove has a 7-day free trial and then a one-time ₹470 Lifetime Pro. Download a model, turn on airplane mode, and see what your own voice does. It is also on Google Play.

Disclosure: I am the developer of Utterwove (Kriva Studio, India). This article was drafted with AI assistance.

來源:DEV Community · dev.to