English | 简体中文

Create a voice conversation

Use the SDK's built-in voice template on your computer: robot microphone → ASR → LLM → TTS → robot speaker. Voice logic belongs to the Application; no ESP32 firmware changes are required, and the robot chip does not run the language model locally.

Prerequisites

Complete SDK installation and pairing. This guide targets SDK 0.1.12. Use the pinned source installation in that guide when using unreleased 0.1.12 features. Check watcherobot app init --help and watcherobot app configure --help before continuing.

The default template uses Volcengine ASR, Qwen LLM, and Volcengine TTS. Enable your services and obtain credentials in the Volcengine console and Alibaba Cloud Bailian console. Providers determine pricing, resource access, and voice availability. Audio and text are sent to the configured services; do not test with recordings you do not want to upload.

Create a project

In a terminal with the SDK environment activated:

watcherobot app init my_voice_app --template voice
cd my_voice_app
watcherobot app configure

Enter ASR, LLM, and TTS credentials. Default ASR/TTS each require app_id and access_token; LLM requires api_key. Input is hidden. The wizard saves and validates configuration locally; it does not connect a robot or verify cloud permissions.

Configure models and prompts

File	Purpose
config/models/asr.toml	Recognition provider and parameters
config/models/llm.toml	Language model, endpoint, and generation parameters
config/models/tts.toml	Speech provider, resource, and voice
credentials/*.toml	Private local credentials; never upload to Git or public apps
prompts/system.md, config/persona.toml	Persona prompt and variables
application/voice.py	Your application composition logic

When changing providers, update both model configuration and credentials. Never put keys in source code, prompts, or parameters. Restart the application after changes. SDK upgrades do not rewrite generated projects.

To update one service, run watcherobot app configure --service llm (also asr or tts). Environment-variable references must be available before the Daemon starts; variables set in a new terminal do not reach an existing Daemon automatically.

Run and verify

watcherobot robot status
watcherobot app check .
watcherobot app run

Keep the connected robot in Desktop Link. Say a sentence, wait for a reply, then ask a follow-up. The default half-duplex flow is speak → answer → listen again, without voice barge-in. Do not run a competing application using the microphone, speaker, or RTC.

Passing app check does not prove cloud access or audible playback. Verify recognized text, model replies, synthesized audio, and physical speaker output separately.

Troubleshooting

• No device: check robot status, Wi-Fi, and Desktop Link; restore the connection first.
• 401/403: check provider, region, credential type, expiry, and resource access. Never post full keys in logs or chat.
• No audio: use the audio primitives to distinguish capture, TTS, and playback failures.
• python app.py fails: use watcherobot app run so the Daemon injects channels.

See the SDK 0.1.12 voice template reference (Chinese) for full configuration and extension APIs. This page was checked against source; no paid service or physical voice acceptance was performed on your behalf.