I changed Google’s voice typing with speech recognition that works offline

Google’s voice typing has labored effectively sufficient on my telephone that changing it by no means actually crossed my thoughts. I faucet the microphone, begin speaking, and the phrases normally land on display screen quick sufficient that I barely take into consideration what’s occurring beneath the hood. That ease has made dictation one of many many ways I type on Android, and I’d gotten used to trusting it.

WhisperBoard gave me a purpose to look elsewhere. It runs a speech-recognition mannequin instantly on my telephone, plugs into Android as a daily keyboard, and may transcribe what I say with no need an web connection. That mixture was sufficient to make me marvel how a lot of Google’s voice-typing comfort I might maintain whereas giving myself an explicitly native choice for the precise speech recognition.

Setting up offline dictation takes slightly dedication

A 671MB keyboard shouldn’t be precisely an off-the-cuff obtain

WhisperBoard is not a transcription app that spits out textual content so that you can copy some place else. It’s a full Android keyboard based mostly on HeliBoard, one of many open-source Gboard alternatives I’ve tested. Because of that basis, something I dictate can land instantly in Notes, Messages, a browser, or nearly anyplace else I’d usually kind.

Getting it arrange took slightly extra effort than tapping Google’s microphone icon. WhisperBoard presently ships as an APK by its GitHub releases, so after putting in it I needed to allow the keyboard in Android, swap to it as an enter technique, grant microphone entry, and configure voice enter.

The setting I cared about was Local mode. WhisperBoard 3.8 changed the Whisper fashions utilized in earlier variations with NVIDIA’s Parakeet TDT 0.6B v3 operating by sherpa-onnx. That replace additionally did away with the outdated alternative between Base and Small Whisper fashions and moved to a single multilingual native mannequin.

On my telephone, that mannequin was a 671MB obtain. WhisperBoard’s launch notes put the file at roughly 640MB, so you must price range round 700MB of free storage and use a reliable Wi-Fi connection earlier than beginning.

That’s significantly extra baggage than putting in one other light-weight keyboard, however it buys one thing tangible. Once downloaded, Local mode performs transcription on the telephone with out requiring web entry. WhisperBoard additionally features a Deepgram-powered Cloud mode and an Auto choice that makes use of cloud transcription when a connection is accessible earlier than falling again to the native mannequin. I stored mine on Local as a result of sending the audio elsewhere would have defeated the purpose of testing offline speech recognition within the first place.

WhisperBoard joins a rising crop of options comparable to FUTO Keyboard and its downloadable offline voice models, which makes native voice enter on Android far much less unique than it was once.

Google is quicker, however WhisperBoard does not want the connection

Live transcription spoils you till it’s important to wait once more

Google received the primary comparability reasonably decisively. In Samsung Notes, its voice typing produced phrases on-screen whereas I used to be nonetheless talking. WhisperBoard 3.8 makes use of a push-to-talk workflow as a substitute. I held the microphone button, spoke, launched it, after which waited for the recorded speech to be processed earlier than the textual content appeared. That’s the interplay its present launch is designed round.

That pause turned particularly noticeable with longer sentences. Google has the form of reside, streaming transcription I’ve grown accustomed to after years of utilizing it, whereas WhisperBoard leaves a brief interval between ending a thought and seeing the finished transcription.

I nonetheless discovered the workflow straightforward sufficient to reside with. I might dictate a full conversational paragraph, launch the button, and have the keyboard drop the transcription straight into my observe. There’s a purpose voice-first Android keyboards can become addictive once typing starts to feel optional.

Accuracy held up effectively sufficient to maintain me , even when it wasn’t flawless. One run turned “I’m having fun with” into “I’m a be part of,” whereas one other transcribed primarily the identical phrase appropriately. My higher runs captured almost the whole paragraph precisely, together with contractions and smart sentence breaks, which made the occasional stumble simpler to forgive.

Parakeet helps automated punctuation and capitalization, together with automated language recognition throughout 25 European languages. In apply, I had higher outcomes once I merely spoke usually and let the mannequin deal with the punctuation. Old dictation habits sometimes labored towards me. Saying “full cease” or “exclamation mark” typically put these literal phrases into the doc, even when WhisperBoard had already punctuated the encircling sentence appropriately.

A slower keyboard can nonetheless be the one I would like

I’ll commerce slightly velocity for figuring out the place my voice stays

I would not swap to WhisperBoard as a result of Google abruptly turned unhealthy at voice typing. Google stays quicker in my testing, and supported Pixel telephones can carry out some Google voice-typing processing on-device as effectively. The case for WhisperBoard due to this fact is not that Google invariably ships each spoken phrase to a server.

What I like is how legible the association is. I select Local, the Parakeet mannequin lives on the gadget, and transcription retains working with out a community connection. For anybody already uneasy about how a lot data an Android keyboard can probably deal with, that clearly outlined native path affords some reassurance with out demanding an elaborate privateness routine.

Google wins the race, WhisperBoard wins the useless zone

I’ve already discovered open-source Android replacements that can stand in for Google’s default apps, and WhisperBoard now sits in that very same bucket for me. It asks for slightly additional setup, storage, and endurance in return for a really easy reply to the place the speech recognition is going on.

Google nonetheless wins once I need phrases showing on-screen earlier than I’ve completed the sentence. When I care extra about with the ability to dictate with awful reception or no connection in any respect, I’m tremendous ready just a few additional seconds for my telephone to deal with the job itself.