First Ollama call is slow (cold start)

[lucharo]

Description

The first call to Ollama after starting v2t takes significantly longer than subsequent calls. This appears to be an Ollama cold-start issue.

Logs

2025-12-12 12:08:32.828 | INFO     | voice2text:process_audio:136 - Transcribing...
2025-12-12 12:08:35.247 | INFO     | voice2text:process_audio:141 - Raw: Hey, how is it going? (2.42s)
2025-12-12 12:08:35.248 | INFO     | voice2text:process_audio:147 - Cleaning up...
2025-12-12 12:09:01.533 | INFO     | voice2text:process_audio:165 - Clean: How is it going? (26.28s)  <-- FIRST CALL: 26s
2025-12-12 12:09:04.926 | SUCCESS  | voice2text:process_audio:172 - Pasted!

2025-12-12 12:09:14.448 | INFO     | voice2text:process_audio:136 - Transcribing...
2025-12-12 12:09:16.807 | INFO     | voice2text:process_audio:141 - Raw: Hey, how's it going? (2.36s)
2025-12-12 12:09:16.808 | INFO     | voice2text:process_audio:147 - Cleaning up...
2025-12-12 12:09:17.389 | INFO     | voice2text:process_audio:165 - Clean: How's it going? (0.58s)   <-- SECOND CALL: 0.6s

Analysis

• First Ollama cleanup: 26.28s
• Second Ollama cleanup: 0.58s

This is a ~45x difference. Likely Ollama loading the model into memory on first call.

Possible solutions (for future consideration)

• Warm up Ollama on startup (similar to how we warm up Whisper)
• Document this as expected behavior
• Investigate if ollama has a keep-alive or preload option