Refactor: Isolate concurrency for local vs. remote transcription approaches

[coderabbitai]

Background

Currently, a single lock is used for both local and remote transcription models in the voice recognition agent. This may unnecessarily limit throughput for remote transcription calls that don't require the same locking mechanism as local models.

Discussion Points

• Separate concurrency logic for local vs. remote transcription flows
• Improve throughput for remote transcription calls
• Maintain thread safety for local model usage

References

• Originally discussed in PR: feat: move automatic speech recognition agent to rai_core #357
• Specific comment: feat: move automatic speech recognition agent to rai_core #357 (comment)

Next Steps

• Design a solution that handles concurrency appropriately for both local and remote transcription approaches
• Implement separate flows while maintaining code maintainability
• Measure and compare throughput improvements