Feature Request: Add support for NVIDIA Parakeet (TDT) transcription backend

[Istiyaq-Khan]

Feature Request: Add support for NVIDIA Parakeet (TDT) transcription backend

Problem Statement

Currently, Recordly relies exclusively on whisper.cpp (specifically ggml-small.bin) for local caption generation. While Whisper is accurate across many languages, it suffers from notable issues during desktop screen recordings:

1. Inference Latency on CPU: Generating captions for long clips on standard CPU hardware can be slow.
2. Hallucinations on Silence: Whisper frequently gets stuck in repeating loops or outputs phantom phrases during long gaps of silence or background music.
3. Timestamp Drift: Extracting tight, word-level timestamps for dynamic subtitle animation requires extra passes or heuristic alignments.

Proposed Solution

Add support for NVIDIA Parakeet-TDT (Token-and-Duration Transducer, e.g., Parakeet-TDT 0.6B / 1.1B v3) as an alternative transcription engine in the Captions panel.

Why Parakeet?

• Throughput: Parakeet-TDT runs up to 4x–10x faster than Whisper on consumer hardware.
• Architectural Advantage: Because it uses an RNN-Transducer / TDT architecture rather than an autoregressive encoder-decoder like Whisper, it produces exact token timestamps natively without hallucinating during silent pauses.
• Local Embedded Viability: Parakeet can be packaged and run completely offline without PyTorch/Python using native runtimes like sherpa-onnx (ONNX Runtime bindings for C++/Node/Rust).

Suggested Implementation Approach

• Use sherpa-onnx, which provides lightweight, multi-platform C/C++ and Node.js FFI bindings to run quantized Parakeet INT8 ONNX models offline.
• Add an engine dropdown selector in Captions settings: Whisper (Default) vs. NVIDIA Parakeet (Fast English/European).
• If the Parakeet model weights are not pre-packaged, implement an explicit download hook or allow users to supply a local .onnx model directory.