AI engines (ASR / TTS / LLM)

CrossMeet splits its AI stack into 3 independently swappable engines: ASR (listen), LLM (translate), TTS (speak). This guide covers selection, API keys, local install, and tuning.

1. Architecture

You speak → mic → ASR → source → LLM → translation → TTS → vMic → other side
            ↓                                                ↑
        VAD / segmentation                          Glossary + RAG injected

Each engine is swappable independently. Example combo: local Whisper (ASR) + Claude (LLM) + Aliyun TTS.

2. ASR (speech recognition)

Comparison

Recommended combos

• Daily engineering chat: Whisper-faster local + GPU (privacy + zero cost)
• Important client meetings: Whisper API (cloud, reliable)
• Chinese-heavy: Aliyun ASR + Qwen3 fallback

Configure API key

Settings → AI Engines → engine card → enter API key + Region

Install local Whisper

1. AI Engines → Whisper-faster card → Install Local
2. Pick a model size (match your hardware):
   • base (300 MB) — 8 GB RAM ok, basic quality
   • medium (1.5 GB) — 16 GB RAM, good quality
   • large-v3 (3 GB) — 16 GB+ RAM + GPU 6 GB+, excellent
3. Wait for download (5–15 minutes)
4. Enable GPU: check Use CUDA (NVIDIA driver ≥ 525)

Key parameters

Settings → ASR → Advanced:

Tuning tip: before a meeting, try a sample sentence in settings, dial in until it feels right.

3. LLM (translation model)

Comparison

Recommended

• Business meetings: Claude 3.5 Sonnet (most natural, least "AI translation" feel)
• Budget-conscious: GPT-4o-mini (ultra-low cost, good enough daily)
• Local privacy: Ollama + Llama 3 / Qwen2 70B (needs 64 GB+ RAM or GPU)

Custom prompt templates

Settings → LLM → Prompt Templates:

You are a professional translator.
Translate from {source_lang} to {target_lang}.

Glossary (MUST follow):
{glossary}

Reference (use if relevant):
{rag_chunks}

Context (previous 3 sentences):
{history}

Sentence to translate:
{text}

Output ONLY the translation, no explanation:

Variables:

• {source_lang} / {target_lang} — language codes
• {glossary} — active glossary terms
• {rag_chunks} — RAG-retrieved chunks
• {history} — last N sentences
• {text} — sentence to translate

4. TTS (speech synthesis)

See Virtual microphone.

Key parameters

Settings → TTS → Advanced:

Multi-voice

Configure different voices per language / persona:

• English (primary): OpenAI TTS — alloy (neutral, professional)
• Chinese (primary): Aliyun TTS — xiaoyun (warm female)
• Japanese (secondary): Edge TTS — ja-JP-NanamiNeural

Settings → TTS → Multi-Voice to configure.

5. Hot-swap shortcuts

The top dropdown in the main UI:

• Ctrl+1 — switch ASR to engine slot 1 (defined in Settings)
• Ctrl+2 — switch LLM to engine slot 2
• Ctrl+3 — switch TTS to engine slot 3

Useful when a meeting suddenly turns sensitive and you want to jump to local engines.

6. Where to get API keys

7. Self-hosted / private endpoints

CrossMeet is OpenAI-API-compatible — point it at any private endpoint:

Settings → LLM → Custom Endpoint
Base URL: https://your-proxy.com/v1
API Key: your-internal-key
Model: claude-3-5-sonnet (any name — your proxy decides)

Works with: corporate proxies, local Ollama, One-API / LiteLLM gateways.

8. LAN API sharing

Expose local engines to other machines on your LAN:

1. Settings → Networking → LAN API
2. Enable → pick which engines to expose (ASR / TTS / LLM)
3. Set an access token
4. On other machines, configure CrossMeet:
   • Base URL: http://your-ip:7860/v1
   • API Key: your token

Great for: one GPU workstation serving 10 thin-laptop teammates.

9. Troubleshooting

More: Troubleshooting.

10. Related

• Getting started
• Glossary
• Knowledge base (RAG)
• Virtual microphone