Your voice. Your phone. Your words.

Near real-time transcription. On your Android phone.
100% offline. 100% private. 100% free.

Turn a passing thought into a message. A voice note into text. All on your device, with nothing sent to a speech-recognition server.

Get Utterlane See how it works

Free & open source · Android 8.0+ · ARM64

Illustrative demo
  • On-device recognition
  • 25 languages
  • No subscription
Less typing. More freedom.

Thought → voice → words

Experience the fastest accurate offline
transcription on Android.

Keep up with your thoughts.
Your language. Your device. Your words.

Accurate.Speech with its meaning intact.
Dependable.No connection to depend on.
25 languages.Automatically understood.

Robust by design. Your progress, saved.

Completed transcript segments are saved as you go. If transcription crashes, recovery can preserve the words already processed, while recording history keeps saved audio available to replay or transcribe again.

Recording history defaults to one-hour retention. Keep recordings longer or turn history off in Settings.

Less tapping. More thought.

Your voice moves faster
than your thumbs.

A small keyboard should not set the pace of a big idea. Speak naturally. Utterlane turns your speech into text on your phone, while you keep talking.

≈4×

the pace of smartphone typing, when speaking naturally.

Picture a 100-word message.

Smaller is faster and better.

Thumb typing≈2m 46s
Speaking≈38s

36.2 words/minute typing · about 160 words/minute speaking.
Illustrative estimate from separate studies, before recognition finalization or editing.

The speech advantage comes from human conversation research, not a voice-typing test. Utterlane outputs completed segments; actual turnaround depends on your phone, model, and pauses. Sources & how this comparison works ↓

A closer lookWhat makes Utterlane different?Compare phone keyboards, cloud transcription, offline tools, and Utterlane. Speed Privacy Meetings 8 illustrated comparisonsThese three topics, plus five more.Expand the comparisonHide comparison

Same thought.
A different experience.

Four ways to get your words down on Android. “Other offline transcription” compares the legacy, older-model batch workflows illustrated here. Cloud features depend on the service and plan.

StrengthDepends / limitedLimitation

01 / Writing speed

Keep the thought moving.

Voice removes the tapping bottleneck. Utterlane starts processing while you speak, so the whole recording does not have to wait until the end.

Phone keyboard
Limitation: One tap at a time

36.2 WPM average in the mobile typing study. Speed can be improved by training with sliding on optimized layouts (eg, KASROZ) but retraining can be long and tedious.

Cloud transcription
Strength: Live text is possible

Depends on the service and connection.

Other offline transcription
Depends / limited: Record. Stop. Wait.

Batch workflows process the recording afterward.

Utterlane
Strength: Write as fast as you speak

Clever optimizations allows SOTA models to stream write your thoughts as you speak. Even speakers labelling in meetings transcripts is done in near-real-time.

Near real-time is device- and model-dependent. It does not mean instant word-by-word text or zero finishing time.

02 / Privacy

Your thoughts stay yours.

A private idea does not need a round trip to a speech-recognition server. The audio-to-text journey happens inside your phone.

Phone keyboard
Strength: No speech upload

Keyboard and receiving-app privacy still apply.

Cloud transcription
Limitation: Audio leaves the phone

Processed on a server; policies vary.

Other offline transcription
Strength: Recognition stays local

When using an entirely offline workflow.

Utterlane
Strength: On-device recognition

No speech uploads. No analytics. No account.

03 / Offline reliability

No signal. Still your words.

On a train, in a basement, or with airplane mode on: recognition remains available after model setup.

Phone keyboard
Strength: Keep typing

Text entry itself can work offline.

Cloud transcription
Limitation: Needs a connection

For the cloud recognition step.

Other offline transcription
Strength: Works offline

Once the required model is installed.

Utterlane
Strength: Works offline

Download once, or import a local model.

04 / Recovery

A second chance, built in.

If recognition is interrupted, saved audio lets you try again. Completed transcript segments are saved along the way.

Phone keyboard
Limitation: No recording

Draft recovery entirely depends on the receiving app.

Cloud transcription
Depends / limited: Service-dependent

Recording retention and recovery vary.

Other offline transcription
Limitation: Recovery is not assured

Without saved audio and partial transcripts, a failed session can mean starting over.

Utterlane
Strength: Replay. Retranscribe.

Local audio history plus saved text segments.

One-hour history retention by default; adjustable or disabled in Settings. Recovery covers saved material, not an unconditional guarantee against all failures.

05 / Accuracy

Speech recognition in real environments.

SOTA speech models with clever optimizations battletested for everyday recordings. Speak, whisper, sing, at home or in a noisy train: your words need to follow you anywhere you go, not just in a perfectly soundproofed studio.

Phone keyboard
Not applicable: You choose the words

Typos and autocorrect still need attention.

Cloud transcription
Strength: Modern frontier models

Strong recognition: the experience we aim to bring on-device.

Other offline transcription
Limitation: Legacy-model limitations

Older models (eg, Whisper, Vosk) can struggle with challenging speech.

Utterlane
Strength: Modern models, locally

Parakeet Ultra plus custom vocabulary corrections.

Accuracy varies with model, voice, microphone, and surroundings. Whispered or sung audio may need corrections. Tip: Clear articulation and speaking closer to the microphone can improve accuracy, especially in noisy environments.

06 / Multilingual

Switch languages, not your flow.

Start in one language and continue in another. Automatic language detection supports multilingual recordings without restarting.

Phone keyboard
Depends / limited: Keyboard-dependent

Layouts and multilingual suggestions vary.

Cloud transcription
Strength: Often multilingual

Automatic detection and mixing vary.

Other offline transcription
Limitation: Mixed-language limitations

Legacy workflows can miss mid-segment language switches. Automatic detection and unintended translation depend on the model and settings.

Utterlane
Strength: 25 languages

Automatic language switching detection in the same recording: start in one language, switch to another one, all faithfully recorded in the language spoke in.

Language switching within a recording is supported; reliability of switching inside a single short segment has not been established here.

07 / Meetings

Follow the people. Keep the conversation.

Turn a meeting into a running transcript, with optional streaming speaker labelling to help follow who said what.

Phone keyboard
Limitation: Manual notes

No automatic conversation capture or labels.

Cloud transcription
Strength: Available in services

Meeting transcription can include speaker labels.

Other offline transcription
Limitation: Never live

In these legacy batch workflows, speaker labelling is a post-processing step after the whole recording ends, rather than a live stream.

Utterlane
Strength: Live stream offline speakers labels

Optional automatic diarization (speakers labelling) of up to 8 speakers, in near real-time. Never lose track of a meeting to write down the minutes: let your phone be your scribe.

Speaker labels require a separate model and add processing work. Uncertain speech can appear as “Unknown speaker”.

08 / Ownership and rate limits

More words. No subscription.

A free, open-source tool on the phone you own. Choose your model and your retention settings, without an account.

Phone keyboard
Strength: Unlimited

The good old trusty input scheme, no rates, no limits, except your hands speed.

Cloud transcription
Limitation: Plans and limits vary

Limited free tiers, subscriptions with rate limits, or usage billing.

Other offline transcription
Depends / limited: Technically unlimited, practically not

Can generate unlimited transcripts, but speed, accuracy, or reliability can limit practical use.

Utterlane
Strength: Truly unlimited and yours

Free & open source & accurate & fast enough to hold the promise of unlimited transcripts.

Unlimited means no service-imposed transcription quotas or per-minute charges.

Your voice. In the apps you already use.

Keyboard mic, accessibility button, or floating mic. Custom word corrections for the names that matter to you.

Explore Utterlane
Sources & comparison notes
  1. Palin et al., MobileHCI 2019: 37,370 volunteers averaged 36.2 WPM copying English sentences on mobile devices. Typing WPM uses five-character words; volunteers were self-selected.
  2. Yuan, Liberman & Cieri, Interspeech 2006: spontaneous conversation, not dictation. English Switchboard turn-wise rate: 164 WPM; total-conversation rate: 196 WPM. This illustration uses a rounded 160 WPM speaking pace. Natural spoken words and standardized typing words are different units: the ≈4× comparison is approximate, not a controlled head-to-head result or measured app speedup.
  3. Parakeet Ultra model card: developer-reported improvements over original Parakeet on multilingual, noisy, and long-form audio. These GPU-runtime measurements are not Utterlane phone accuracy measurements.
  4. Utterlane user guide: segment-based output, automatic recognition across 25 languages, history controls, optional speaker labels and their model requirement.

01 / A shorter path from thought to text

A small journey.
Entirely on your phone.

Speak naturally. Utterlane turns your speech into completed text segments while you keep talking. Your phone does the work.

  1. 01Capture your voice
  2. 02Process on-device
  3. 03Turn words into action

Recognition works offline after model setup. Output arrives in segments; timing depends on your device, model, and pauses.

Illustration: voice is captured and processed within your phone, then returned as text.

02 / Ideas, study, and work

For your
everyday needs.

Keep a thought. Read a recording. Follow a meeting. All on your device.

Illustration: a person captures a bright idea as a private note on their phone.

Capture ideas that stay yours.

Speak a passing thought, a research idea, or your next big plan. Turn it into a note on your device, without handing it to a cloud transcription service where it could be scooped.

Illustration: a student turns a recorded lecture into written notes to study.

Turn recordings into something you can read.

A voice message, a recorded lecture, or audio saved from a video: share it to Utterlane and get the words in writing. Revisit your professor’s explanations or read a message when listening is inconvenient.

Illustration: three meeting participants’ speech bubbles flow into a phone, building a transcript with a matching label for each speaker.

Follow the meeting, keep the words.

Get a running transcript while the conversation continues, with speaker labels to help you follow who said what. Text arrives in completed segments, right on your device, keeping the discussion local.

25 languages.Automatically detected. Mix languages naturally.
Your own vocabulary.Custom word corrections.
More ways to speak.Keyboard, accessibility, floating mic.

03 / Local by design

Your words are yours.
Let’s keep it that way.

A thought does not need to travel to a server to become text. Utterlane keeps speech recognition where it belongs: on your device.

Read the privacy policy

No speech uploads.

Your audio and transcripts are not sent to a speech service. Recognition runs locally, even in airplane mode.

No account. No analytics.

There is nothing to sign up for. Recording history stays on your device, with one-hour retention by default. You control what you keep or export.

Open source. Open to scrutiny.

The code is available under Apache-2.0. Read it, build it, improve it. Good tools should be accessible to everyone.

04 / A little setup. A lot of possibility.

Quickstart

  1. 01

    Get the app.

    Choose an Utterlane APK from GitHub Releases when available, or build it from source.

    View releases
  2. 02

    Choose your model.

    Download a speech model in the app, or import existing model files. Utterlane supports the latest state-of-the-art speech-to-text models, including Moondream Parakeet Ultra models and any model supported by the CrispASR backend, such as R2T2, MegaASR, and Qwen3. Recognition works offline from then on.

    Size varies by model
  3. 03

    Tap. Speak. Done.

    Allow microphone access and enable your preferred input method. Focus a text field, tap the mic, and speak.

    Already have a recording? Share it from another app to Utterlane to transcribe it directly.

    See the setup guide
Android 8.0 or laterARM64 deviceSpace for a speech modelPrefer to build it yourself?

A few good questions

FAQ

A little clarity before
your first dictation.

Ask on GitHub
Does it really work offline?

Yes. Once a speech model is installed, recognition works without a network connection. The initial model download uses Hugging Face; you can also import model files from a local folder.

What does “near real-time” mean?

Completed text segments arrive while microphone capture continues. This is not instantaneous word-by-word output. Timing depends on your phone, the selected model, and pauses in your speech.

The default Parakeet v3 backend simulates streaming over an offline model. Read how streaming works.

Can I dictate into other apps?

Use a compatible keyboard’s microphone button, the accessibility button, or the floating mic. Utterlane inserts text into the focused field and can fall back to clipboard or transcript recovery when insertion is unavailable.

Integration depends on the receiving app and keyboard. HeliBoard is one compatible keyboard; AOSP Keyboard does not offer this voice-input integration.

Which languages and audio files are supported?

Utterlane automatically detects 25 recognition languages: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish, and Ukrainian.

Share or open OPUS, AAC, OGG, M4A, MP3, and WAV files, subject to your device’s codec support. You can also monitor selected folders for newly saved audio.

What happens to my recordings?

Microphone history defaults to one-hour retention, so you can replay or retranscribe saved recordings. Change retention or disable history in Settings; existing choices are preserved. Saved history is private to the app and excluded from Android cloud backup and device transfer.

Completed transcript segments are saved as they arrive, and interrupted recordings can be recovered from saved audio. A segment still being processed may need to be transcribed again. Temporary text files support long transcripts and recovery. When you choose to copy, share, or export, the receiving app has its own privacy behavior.

Is it really free?

Yes. No account or subscription is required. Utterlane is open source under Apache-2.0; models and third-party libraries retain their own licenses.

If it helps you, help it grow: report a bug, improve a translation, or contribute code. Contributions are welcome.

Built with care. Free as in free speech (and beer!).

Your voice deserves
this much freedom.

No account. No subscription. Open-source speech recognition that runs on the phone you already own.

Get Utterlane
Offline after setup. Open source from the start.