---
type: "post"
id: "phone-whisper-simple-voice-transcription-for-android-phones-4caa9a"
slug: "phone-whisper-simple-voice-transcription-for-android-phones-4caa9a"
title: "phone-whisper: Simple voice transcription for android phones with Whisper and local models"
description: "phone-whisper is an open-source Android app for entering text through voice dictation without changing keyboards."
retrieval_nugget: "phone-whisper is an open-source Android app for entering text through voice dictation without changing keyboards. It can transcribe locally with sherpa-onnx or use OpenAI Whisper in the cloud, while an optional OpenAI pass cleans punctuation and grammar. Its recent 1.3.0 work focuses on keeping the service alive under aggressive Android power management."
published_at: "2026-10-31"
updated_at: "2026-10-31"
record_date: "2026-10-31"
date_kind: "published_at"
topics: ["local-models"]
entities: ["phone-whisper","sherpa-onnx","OpenAI Whisper","OpenAI","Android"]
related_patterns: []
novelty_label: "new_fact"
what_changed: "Phone-whisper adds an Android dictation app that lets users dictate without switching keyboards; the prior voice-dictation stories cover Wispr Flow."
prior_refs: ["memory:6e62f35a-2674-89f2-a263-b5cf11bc604c","memory:3e999168-2119-8c22-af16-13237353ad8a"]
source_url: "https://github.com/evilknivel/phone-whisper"
source_urls: ["https://github.com/evilknivel/phone-whisper"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:phone-whisper-simple-voice-transcription-for-android-phones-4caa9a"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/phone-whisper-simple-voice-transcription-for-android-phones-4caa9a/","markdown":"https://newruntime.com/posts/phone-whisper-simple-voice-transcription-for-android-phones-4caa9a.md","json":"https://newruntime.com/posts/phone-whisper-simple-voice-transcription-for-android-phones-4caa9a.json"}
---

# phone-whisper: Simple voice transcription for android phones with Whisper and local models

## Retrieval answer

phone-whisper is an open-source Android app for entering text through voice dictation without changing keyboards. It can transcribe locally with sherpa-onnx or use OpenAI Whisper in the cloud, while an optional OpenAI pass cleans punctuation and grammar. Its recent 1.3.0 work focuses on keeping the service alive under aggressive Android power management.

## Summary

phone-whisper is an open-source Android app for entering text through voice dictation without changing keyboards. It can transcribe locally with sherpa-onnx or use OpenAI Whisper in the cloud, while an optional OpenAI pass cleans punctuation and grammar. Its recent 1.3.0 work focuses on keeping the service alive under aggressive Android power management.

## What happened

The GitHub repository for [phone-whisper](https://github.com/evilknivel/phone-whisper) describes a simple voice-transcription app for Android phones. Its intended interaction is narrow but practical: dictate text into the phone without switching between keyboards. The project is written in Kotlin and is open source under the Apache 2.0 licence named in its repository materials.

The app offers two transcription paths. One is local transcription using a sherpa-onnx model. The other is cloud transcription through OpenAI Whisper. The repository also describes an optional post-processing step through OpenAI that can correct punctuation and grammar after transcription. Those choices matter because they separate the initial speech-to-text operation from a later text-cleanup operation, rather than presenting the result as one indivisible voice feature.

The repository materials also indicate that local transcription is packaged as part of the Android build rather than merely described as a future integration. A recent build-related commit says native sherpa-onnx and ONNX Runtime libraries were checked into the project so that a plain Gradle build includes local transcription support. That is a meaningful implementation detail, although it does not establish the size of the resulting application, the models a user receives, or how a user selects among them.

The project has been actively updated. The repository shows 53 commits, six branches, and five tags at the captured time. Its latest visible commit is dated October 28, 2026 and is associated with version 1.3.0. That change adds a “Keep running” setup path intended to deal with phone-maker settings that terminate background services. The project is therefore addressing not only transcription quality or output formatting, but the operating conditions required for a dictation utility to remain present when a person wants to use it.

The accompanying commit description is unusually specific about the problem. On a Xiaomi phone running HyperOS, clearing recent apps can kill the process; without Autostart enabled, Android may not restart it, leaving the service listed as crashed until the relevant accessibility setting is toggled. The project says it verified that enabling Autostart caused the service to be rebound within five seconds on that device. The new settings assistant detects the manufacturer and points users toward relevant controls, including Xiaomi Autostart and battery settings, Samsung never-sleeping apps, Huawei and Honor app-launch controls, ColorOS auto-launch, vivo background start, ASUS, Meizu, Transsion, or general battery-optimization settings.

The assistant attempts known system settings screens in sequence and falls back to app information when necessary. The source says the setup is also exposed as a new assistant step and as “Keep running” on the main screen. This design acknowledges that device-specific configuration is part of deploying the feature, not merely a troubleshooting note for a minority of users.

This is not presented as a new Android platform capability or as a broadly distributed commercial product. It is a repository-level release from a small project. The source does not state a price, a Play Store distribution path, model-download sizes, supported Android versions, transcription accuracy, latency, or a complete language list. It does, however, show a concrete attempt to make a local-or-cloud dictation utility survive the operating-system constraints that often determine whether such a tool is usable in practice.

## What the archive adds

We have not covered this before; it opens a new thread.

## Trend signal

No supplied pattern or archive assumption genuinely fits this small Android dictation release.

## Why it matters

For teams building AI-enabled mobile tools, the useful signal is not that speech recognition exists on a phone. It is that product reliability sits above model selection. phone-whisper combines a local inference option, a cloud option, and optional language cleanup, but its latest release spends visible effort on service survival, manufacturer-specific settings, and recovery after a process is killed.

That is a familiar mobile constraint with particular force for AI features that are expected to be available at the moment of input. A technically capable transcription stack is of limited value if a user must reopen or reconfigure it after the operating system clears background work. Builders evaluating voice input should therefore treat lifecycle behavior, power-management exceptions, and setup clarity as part of the feature—not implementation details to solve after choosing a model.

The app also illustrates a useful product decomposition. Transcription, cloud fallback, and text cleanup can be separate choices with different operational and privacy implications. A team does not need to treat voice input as a single vendor decision. But the source makes clear that exposing those choices does not remove the need to test whether the input surface is actually alive when the user reaches for it.

## Limits and uncertainty

The repository is the only supplied primary source, and it does not provide independent tests of recognition quality, privacy properties, battery use, or background reliability across devices. The five-second rebinding observation is explicitly tied to a Xiaomi HyperOS device, not a general Android result. “Local” transcription indicates an on-device path, but the supplied material does not establish what data remains on-device in every configuration or how models are obtained and updated.

The source also does not give installation instructions in the supplied evidence, pricing for any optional OpenAI use, a stable release-channel description, or comparative results against Android’s built-in dictation. The project’s current activity and version number should not be read as evidence of broad adoption or production support.

## Skill to build now

If a team is adding voice input to an Android workflow, test the lifecycle before polishing prompts or transcript cleanup. On at least one aggressively managed device, run the intended dictation flow, clear recent apps, enable and disable the relevant battery or autostart exception, and record whether the service returns without user intervention. Make that recovery path visible in onboarding rather than burying it in a support article.

## Reader takeaway

phone-whisper is a small open-source example of Android voice dictation with local sherpa-onnx transcription, optional OpenAI Whisper, and optional text cleanup. Its latest work highlights the harder product question: a voice feature must remain available after Android and device-maker power controls try to stop it.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Local Models](https://newruntime.com/topics/local-models.md): Explore the local-models topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)