BiBi-Keyboard

Project Overview

BiBi Keyboard is a Kotlin-built Android voice keyboard: open the keyboard, hold the mic and speak naturally; release to recognize, with silence auto-stopping the recording and an LLM polishing the transcript afterwards.

It supports 18 ASR engines, 12 cloud services and 6 local models (SenseVoice, FunASR Nano, Qwen3-ASR, Parakeet, FireRedASR, X-ASR) that work fully offline. Backup engines and local fallback keep input reliable on unstable networks.

Item	Value
Platform	Android
Language	Kotlin
ASR engines	18 (12 cloud + 6 local offline)
Post-processing	LLM text polish (OpenAI-compatible and more)
License	Apache-2.0

Hold to talk, release to text, with cloud and local engines as double insurance.

Key Features

• Hold to talk: hold the mic to record, release to recognize, dead simple
• Smart stop: recording ends automatically on silence, no manual tapping
• Multi-engine: 18 ASR providers, switch between cloud and local freely
• Offline recognition: 6 local models including SenseVoice and Qwen3-ASR work with no network
• Backup and fallback: configurable secondary engines and local fallback for weak networks
• AI post-processing: an LLM trims filler words and fixes punctuation for cleaner output
• History management: cached audio in history can be re-run through recognition and post-processing
• Guided setup: first-run onboarding walks through input method and permission setup

Target Audience

• Users who speak faster than they type
• Commuters and drivers needing hands-free input
• Privacy-minded users wanting offline recognition
• Heavy voice users needing mixed-language input and engine redundancy

Quick Start

1. Download the APK for your architecture from Releases and install
2. Follow the guide to enable and switch to this input method in system settings
3. Grant microphone and network permissions
4. Pick a cloud engine and enter its API key, or download a local model for offline use
5. Open any text field, hold the mic and start talking

Core Advantages

• Widest engine choice: 18 ASR engines one tap away, failover is instant
• Works offline: local models free voice input from network dependence
• Cleaner transcripts: LLM post-processing turns speech into polished text
• Open source: Apache-2.0 licensed, code fully public

System Requirements

Item	Requirement
OS	Android
Architectures	arm64-v8a, armeabi-v7a
Permissions	Microphone, network (cloud engines)

Notes

• Bring your own keys: cloud ASR services require API keys you register yourself
• Local models download separately: offline models are downloaded in-app and take some storage
• Sensitive permissions: keyboards can read typed content, get installers only from official Releases