PKU harnessVoice input
Speak instead of typing in the chat box. Set up a speech-to-text service with your own key (Alibaba Cloud Bailian, the default, where text appears while you talk; DeepSeek, with the key you already use for the AI; or SiliconFlow) and transcribe audio files.
With voice input you tap the microphone in the chat box and speak. Your words turn into text in the message box, and you can edit them before you send. Like the AI model, speech recognition comes from a provider you choose, with your own API key.
Your recordings go straight from the phone to that provider. They are deleted from the phone once transcribed, and PKU harness never receives them.
The services you can choose
Open Settings → Voice input. The screen lists the services provided by plugins, with the default first.
| Service | How text appears | Models | Where to get a key |
|---|---|---|---|
| Alibaba Cloud Bailian (live) (default) | While you talk | Fun-ASR (default, ¥0.00033 per second), Paraformer v2 (¥0.00024 per second) | The Bailian console, Beijing region |
| DeepSeek | In phrases, at each pause | DeepSeek speech recognition (detects the language: Mandarin, Cantonese, English and more), $0.116 per hour of audio, currently 50% off at $0.058, billed per second | No separate key: it uses your DeepSeek key from Settings → AI provider |
| SiliconFlow (硅基流动) | In phrases, at each pause | SenseVoice (Mandarin, English, Cantonese, Japanese, Korean), TeleSpeech (better on Chinese dialects) | cloud.siliconflow.cn |
| OpenAI-compatible | In phrases | You type the model ID | Any service with an /audio/transcriptions endpoint |
The prices are the ones the app shows next to each model; the provider's price list is what counts.
We recommend Alibaba Cloud Bailian with Fun-ASR: the text appears as you speak, and in our tests it recognised campus place names best. If you already use Bailian for the AI model, the same key works here.
If you use DeepSeek for the AI model and don't want another account, choose DeepSeek. It needs no extra key. DeepSeek has no live recognition, so the app sends your speech in pieces cut at each pause, and the text appears phrase by phrase rather than word by word. Prices are from DeepSeek's price list.
Set up Alibaba Cloud Bailian
- Create a Bailian API key in the China North 2 (Beijing) region. The steps are the same as in AI provider → Get an Alibaba Cloud Bailian API key. A key from another region doesn't work.
- In the app, open Settings → Voice input and tap Alibaba Cloud Bailian (live).
- Under Model, tap Fun-ASR.
- Paste the key into API Key. The Get an API key button opens the Bailian key page.
- Tap Save.
The address is fixed (wss://dashscope.aliyuncs.com/api-ws/v1/inference) and shown under the model
list.
Set up DeepSeek
- Make sure your DeepSeek API key is saved in Settings → AI provider. If it isn't, follow AI provider → Get a DeepSeek API key.
- In the app, open Settings → Voice input and tap DeepSeek.
- The screen shows "Uses your DeepSeek API key from AI provider settings. No separate key needed." There is nothing to paste. If it says "Add your DeepSeek API key in AI provider settings first." instead, tap Open AI provider settings and save the key there.
- Tap Save.
Set up SiliconFlow
- Sign up at cloud.siliconflow.cn and complete real-name verification if it asks.
- Open API keys (API 密钥), create a key and copy it.
- In the app, open Settings → Voice input and tap SiliconFlow.
- Under Model, tap SenseVoice, or TeleSpeech if you speak a Chinese dialect.
- Paste the key into API Key and tap Save.
Speak a message
- On the Chat tab, tap the microphone in the message box. The first time, Android asks for microphone access; allow it.
- Speak. The box shows Listening… and your words.
- When you finish, tap the check button (Done, transcribe). The text stays in the message box.
- Edit it if you like, and send.
To throw the recording away, tap the close button (Discard recording).
Transcribe an audio file
You can also attach an audio file in chat: tap +, then Files, and choose a recording (up to 50 MB). The voice input service turns it into text, and the assistant reads the transcript.
Alibaba Cloud Bailian (live) only transcribes live speech. For audio files, choose SiliconFlow or an OpenAI-compatible service in Settings → Voice input.
DeepSeek can transcribe audio files too, but it only accepts WAV or Ogg Opus recordings of up to 120 seconds and 8 MiB. Longer recordings and MP3 or M4A files may fail, so we recommend SiliconFlow for audio files.
When voice input doesn't work
| Message | What to do |
|---|---|
| Set up voice input first (Settings → Voice input). | Choose a service and save it |
| Add an API key for voice input first (Settings → Voice input). | Paste a key and tap Save |
| Microphone access is off… | Allow the microphone for PKU harness in Android's app settings |
| Nothing was heard… | Speak again, closer to the phone |
| Voice input failed: … | The reason comes from the provider. For Bailian, check that the key is from the Beijing region and the account has a balance |
You can also ask the assistant to switch the voice input service, for example "Use SiliconFlow for voice input". It can't enter the key; it opens this screen for you.
AI provider
Set up the AI model first — which provider to choose in mainland China (DeepSeek, Alibaba Cloud Bailian, Baidu Qianfan), how to get an API key step by step, where to enter it, the account balance, the context budget and costs.
Campus Card and payment-code sharing
Your Campus Card balance, payment code, top-ups, network fee and bath card in the Campus wallet plugin, and lending your payment code to a friend with Share payment code.