Version 1.1.0

Release Date: 27 November, 2025

• Enhanced Restart Flow After New Voice Installation
  Fixed issues causing delayed or incorrect restarts when adding new voices. The system now restarts cleanly and reliably after each addition.
• Smarter Auto-Language Switching
  Improved detection and switching logic for multilingual content, resulting in smoother transitions and fewer misclassifications.

Rehnuma Manager UI Upgrade

• Polished interface for better usability and workflow.
• Added support for registering other Piper-based English voices/models.
• Introduced quick voice selection directly inside the Rehnuma Voice Manager panel.

Voice Improvements

• Rehnuma Akber: 

• New Voice: Rehnuma Arfa (Female): 

Miscellaneous

• Various stability fixes, performance tuning, refactoring, and internal improvements.

Version 1.0.4

Release Date: 30 October, 2025

• Enabled ONNX Runtime graph optimizations
• Set GraphOptimizationLevel to Level3 and enabled memory pattern reuse for inference sessions

• Impact: 15–45% faster model inference; fewer allocations during repeated calls

• Added phonemization result caching

• Thread-safe cache keyed by voice and processed text (includes Arabic diacritized text)

• Impact: 50–90% lower latency for repeated or similar utterances; lower CPU usage

• Fixed exponential growth in real-time streaming chunk size

• Replaced multiplicative chunk size update with constant base chunk size per stream

• Impact: predictable latency, reduced memory usage, prevents performance degradation over time

• Enabled GPU acceleration on Windows via DirectML

• ONNX Runtime provider ordering: DirectML (preferred) → CPU fallback

• Impact: large speedups on compatible GPUs without configuration changes

• Tuned real-time streaming parameters

• Updated defaults: chunk_size=64, chunk_padding=2

• Impact: lower initial latency while maintaining stream smoothness

• Increased gRPC channel buffer sizes

• mpsc channel capacity raised from 512 to 1024 for both synthesis endpoints

• Impact: reduced backpressure and blocking under bursty load; smoother streaming

• Configured gRPC server for high concurrency

• Set concurrency_limit_per_connection=1024, max_concurrent_streams=1000, tcp_keepalive=60s, tcp_nodelay=true

• Impact: improved throughput and responsiveness under concurrent client load

• Right-sized synthesis thread pool

• Reduced from num_cpus4 to min(num_cpus2, 16), with a floor of 2

• Impact: less context switching, more stable latency, better CPU utilization