Voice Input Basics

The core of BiBi Keyboard is high-quality speech recognition. It supports multiple ASR engines and recognition modes, so you can use voice input in any app.

How It Works

Voice input has three stages:

1. Recording: the app records your voice. Depending on settings, it can auto-stop on silence or stop manually.
2. Recognition: audio is sent to an ASR engine (cloud or local) and transcribed into text.
3. Output: the transcript can optionally be refined by AI post-processing, then inserted into the current editor.

Supported ASR Providers

BiBi Keyboard supports 12 ASR providers, grouped into cloud and local:

Cloud ASR

Provider	Streaming	Duration limit (non-streaming)	Notes
Volcengine	✅	1 hour	New users often get free quota; supports bidirectional streaming
SiliconFlow	❌	20 min	Built-in free ASR (SenseVoiceSmall / TeleSpeechASR); supports Qwen3-Omni transcription (own key)
ElevenLabs	✅	20 min	High-accuracy English; supports both file and streaming
OpenAI	❌	20 min	Default gpt-4o-mini-transcribe; you can use any compatible Audio Transcriptions model
DashScope (Alibaba)	✅	3 min	qwen3-asr-flash; supports streaming and non-streaming
Gemini (Google)	❌	4 hours	File-based multimodal speech understanding
Soniox	✅	1 hour	Supports multi-language prompts; both streaming and file modes
Zhipu (GLM)	❌	20 min	GLM-ASR; supports context prompt parameters

Local ASR (Offline)

Provider	Streaming	Duration limit (non-streaming)	Notes
SenseVoice	Pseudo ¹	5 min	Based on sherpa-onnx; multilingual
TeleSpeech	Pseudo ¹	5 min	Based on sherpa-onnx; optimized for CN
Paraformer	✅	Unlimited ²	Pure local streaming recognition
Zipformer	✅	Unlimited ²	Pure local streaming recognition

Notes

¹ Pseudo-streaming: shows partial results based on VAD segmentation, but it is not true real-time streaming.

² Streaming mode has no duration limit (continuous recognition).

The "Duration limit (non-streaming)" here is the app's single-segment recording cap used to control segmented recording behavior. It does not represent provider billing limits or total free quota. For example: Volcengine often provides ~20 hours of free quota for new users; SiliconFlow provides a built-in free ASR service with no total duration quota. For other providers, check their consoles for quota/billing.

For more details on supported models, recommended configs and updated quotas, see the Providers & Models Guide.

Cloud vs Local

Advantages of cloud ASR

• Higher accuracy: large cloud models often perform better
• Multilingual: better for code-switching, dialects, and more languages
• No maintenance: no model download; updates are handled by the provider

Advantages of local ASR

• Fully offline: no network required; privacy-friendly
• Lower latency: no network transfer
• No data usage: good for limited networks
• No API quota: no need to worry about API costs and limits

Streaming vs Non-streaming

Streaming recognition

How it works: upload audio while recording, and get partial results in real time.

Pros:

• ✅ real-time feedback
• ✅ no duration limit
• ✅ lower latency

Supported engines:

• Cloud: Volcengine, Soniox, DashScope, ElevenLabs
• Local: Paraformer, Zipformer

Non-streaming recognition (file upload)

How it works: upload the whole audio file after recording stops.

Pros:

• ✅ potentially higher quality (global analysis on full audio)
• ✅ simpler and stable
• ✅ supports more providers

Cons:

• ⚠️ duration limit (see tables above)
• ⚠️ recognition starts only after recording stops

Suggestions

• For providers that support both modes, switch under Settings → ASR Settings → [Provider Settings].
• Streaming is great for long recordings and low-latency feedback.
• File mode is great for short audio when accuracy matters more.

Segmented Recording

For non-streaming engines, if a recording exceeds the app's single-segment limit for that provider, BiBi Keyboard automatically performs segmented recording.

How it works

1. Auto split: near the limit, the current segment is cut and a new segment starts
2. Background upload: segments are uploaded/recognized in background while recording continues
3. Seamless UX: UI stays in recording state without noticeable interruption
4. Merge results: transcripts from segments are concatenated automatically

Per-provider segment limits (app-side cap)

Provider	Segment cap	Notes
Volcengine	1 hour	Official max per request is ~2h; app uses 1h as a safety margin
SiliconFlow	20 min	App default; unrelated to billing/quota
ElevenLabs	20 min	App default to avoid failures on very long audio
OpenAI	20 min	App default; tune model/usage as needed
DashScope	3 min	Default qwen3-asr-flash; app segment cap is 3 min
Gemini	4 hours	Official max is ~9.5h; app uses 4h as a safety margin
Soniox	1 hour	No strict official max found; app defaults to 1h
SenseVoice	5 min	Local performance cap to avoid excessive RAM/time
TeleSpeech	5 min	Local performance cap to avoid excessive RAM/time

Notes

• Streaming engines (Paraformer, Zipformer, etc.) have no duration limit.
• Segmented recording works only in non-streaming mode.
• Each segment may incur a separate API call cost (for cloud providers).

One-tap Setup Options

1. Use the free service (recommended for beginners)

No config needed:

1. Open the app; it defaults to SiliconFlow free service
2. Under Settings → ASR Settings → SiliconFlow, switch between the two free models (FunAudioLLM/SenseVoiceSmall and TeleAI/TeleSpeechASR)

2. Configure a cloud provider

Using Volcengine as an example:

1. Create an account in the Volcengine console
2. Create an app and obtain App Key and Access Key
3. In BiBi Keyboard, go to Settings → ASR Settings → Provider and select Volcengine
4. Fill in credentials and save

3. Auto-configure local models

1. Download SenseVoice Small from the model release
2. Extract to Android/data/com.brycewg.asrkb/files/sensevoice/
3. The app will automatically select SenseVoice under Settings → ASR Settings → Provider

Tip

First load of local models may take a few seconds. You can enable "Preload model" (SenseVoice / TeleSpeech / Paraformer / Zipformer all support it) so the model is loaded when the keyboard or floating ball is first shown, reducing the first-recognition latency.

Related

• Floating Ball - voice input anywhere
• AI Post-processing - refine transcripts with LLM
• Recording Modes - press-and-hold vs tap-to-toggle
• Auto-stop on Silence (VAD) - stop recording automatically