CrispASR started as a fork of whisper.cpp and extends that base into a single C++ binary -- no Python, no PyTorch, no pip install -- that loads a GGUF model file and can auto-detect which of its many built-in speech backends to run. The README's own headline claims 54 ASR backends and 55 TTS engines, though other sections of the same README cite 52 TTS engines or describe the project as having '100+ backends' in aggregate, so the exact count varies depending on which part of the README is read. Backends span architectures such as OpenAI Whisper, NVIDIA Parakeet and Canary, Mistral Voxtral, Cohere Transcribe, and Qwen3-ASR for transcription, and Kokoro, Qwen3-TTS, VibeVoice, and CosyVoice3 for synthesis. Beyond core ASR/TTS, it ships standalone forced alignment (--align-only) and CTC forced aligners for word-level timestamps, text-to-text translation backends (m2m100, WMT21, MADLAD-400), post-processing models for punctuation restoration, truecasing, and language identification, and music/audio-analysis tasks such as source separation, piano transcription, guitar tablature, beat/chord detection, and pitch estimation. All backends compile to a 4.3 MB WebAssembly build via build-wasm.sh for fully client-side browser use, and a live HuggingFace Space demo is provided. Prebuilt Linux tarballs differ in GPU fallback behavior: the -cuda tarballs fall back to CPU when no compatible GPU is present, while the -hip and -vulkan tarballs require the matching GPU driver and do not fall back. CrispASR is MIT-licensed, the same license as the upstream whisper.cpp it is built on.