Fix: Improve Telugu Live Transcription Reliability (WebSocket + Streaming)

Description

• Update live Telugu transcription pipeline to improve reliability and accuracy without changing the ASR model.
• Currently, the system depends on unstable WebSocket handling and inconsistent audio input, leading to missing final transcripts and noisy partial outputs.
• This change focuses on fixing streaming, session lifecycle, and audio pipeline issues to ensure consistent UI behavior and better transcription quality.

Changes Required

• Ensure client sends end event and waits for final (is_final=true) before closing WebSocket
• Add timeout-based fallback only if final transcript is not received
• Enforce Telugu (te) language across streaming pipeline
• Apply mic constraints (echoCancellation, noiseSuppression, autoGainControl)
• Improve session handling (clean start/reset per recording)
• Refine chunk processing and silence thresholds
• Clean partial vs final transcription behavior via config

Motivation

• Eliminates missing final transcript issue in UI
• Improves stability of real-time transcription
• Enhances accuracy without changing model
• Reduces dependency on audio/environment inconsistencies
• Provides consistent streaming experience across sessions

Acceptance Criteria

• Final transcript always appears in UI after stop
• WebSocket does not close before final message
• Telugu-only transcription is enforced
• Partial and final outputs behave consistently
• Stable results for short Telugu speech