Worked on enhancing real-time text-to-speech streaming in the Blaizzy/mlx-audio repository by implementing chunked audio generation with overlap-add decoding. This approach addressed boundary artifacts between audio chunks and introduced a 16-frame cross-chunk context, improving audio quality during streaming. Developed new parameters for streaming control, allowing fine-tuning of responsiveness and latency for long-form inputs. Leveraged Python and applied expertise in audio processing, machine learning, and streaming technology to achieve substantial reductions in time-to-first-audio across various input lengths. The work focused on enabling practical real-time use cases and optimizing the performance of 4-bit quantized models for streaming scenarios.