Improve performance for ASR

[openminddev]

This pull request introduces several improvements to the ASR (Automatic Speech Recognition) processing pipeline, logging configuration, and websocket server responsiveness. The main focus is on making the ASR model more adaptive to incoming audio sample rates, fine-tuning silence detection thresholds, and improving system responsiveness by reducing sleep intervals in async message processing.

ASR Processing Improvements:

• The ASR processor now waits for the first audio chunk to dynamically set the sample rate, reinitializing the model if the incoming rate differs from the default. This ensures the model matches the actual audio input, improving accuracy and flexibility.
• Default ASR silence detection thresholds and history durations are now set more thoughtfully, with new values for stop_history_eou, stop_threshold_eou, stop_history, and stop_threshold to balance between timely finalization and avoiding mid-sentence cuts.
• The default sample rate for ASR (--asr-sample-rate-hz) is increased from 16kHz to 48kHz, supporting higher-fidelity audio.

Logging Enhancements:

• Logging configuration now uses force=True to ensure that the logging setup overrides any existing configurations, providing more predictable logging behavior.
• ASR logging messages are improved for clarity, distinguishing between final and interim results.

Websocket Server Responsiveness:

• The async websocket server reduces sleep intervals when waiting for messages, improving responsiveness for both global and per-connection message processing. [1] [2]