Skip to main content
Version: V2.0.5.1

3.7.Frequently Asked Questions

Q1: How to modify wake-up words?​

A: Modify in tk_audio_process.py:

self.wake_up_words = ["TienKung", "sky", "space"]  # Change to your wake-up words

Q2: How to increase speech response speed?​

A: Reduce LLM model or use faster model:

# In llm_client.py
self.model = "qwen2.5:0.5b"  # Change to smaller model, but logic generation quality may need verification

Q3: How to adjust TTS speech rate?​

A: Adjust in piper_provider.py:

self.piper_syn_config = SynthesisConfig(
    length_scale=0.8,  # < 1.0 faster, > 1.0 slower
    ...
)

Q4: Does it support multiple languages?​

A: Need to download Piper models for corresponding languages:

# Download language-specific voice models
https://huggingface.co/rhasspy/piper-voices/tree/main

Then modify model path in piper_provider.py.

B: The Funasr on x86 also needs redeployment with models supporting other languages, reference: https://github.com/modelscope/FunASR

Q5: How to implement multi-turn conversation memory?​

A: Adjust history length in llm_client.py:

self.history = deque(maxlen=10)  # Keep latest 5 turns (10 messages)

Q6: Can other Ollama models be used offline?​

A: Yes, but Ollama models must be downloaded in advance:

ollama pull qwen2.5:1.5b

Then modify the model specification when calling llm_client.

Q7: Error when calling funasr_client for speech recognition?​

Go to x86 machine and check if funasr service Docker container is running:

docker ps|grep asr

Q8: How to handle unstable network?​

A: Add retry mechanism:

for attempt in range(3):
    try:
        result = self.asr_service.to_text(audio_bytes)
        return result
    except Exception as e:
        if attempt < 2:
            time.sleep(2 ** attempt)  # Exponential backoff
        else:
            raise