Model support per CrispASR — pure C++ inference with GGUF (no Python/NeMo needed)
#38
by cstr - opened
We've built a complete C++ runtime for Parakeet-TDT in CrispASR, a multi-backend ASR tool based on ggml. One binary, one GGUF file — no Python, no NeMo, no pip install.
What works:
- Full pipeline: mel → FastConformer encoder → TDT decoder with word timestamps
- Native word-level timestamps from the TDT duration head (no separate aligner needed)
- 25 European languages with automatic detection
- ~1x realtime on CPU (11.3s for 11s audio on a 4-core Xeon)
- Streaming from mic/stdin (
--stream,--mic,--live) - Speaker diarisation, language ID, SRT/VTT/JSON output
- GPU acceleration via CUDA / Metal / Vulkan (ggml backends)
- Punctuation toggle (
-ppto strip punctuation)
Quick start:
git clone https://github.com/CrispStrobe/CrispASR && cd CrispASR
cmake -S . -B build && cmake --build build -j8
# Auto-download and transcribe
./build/bin/crispasr --backend parakeet -m auto -f audio.wav
# With word-level SRT output
./build/bin/crispasr -m parakeet-tdt-0.6b-v3.gguf -f audio.wav -osrt -ml 1
Pre-built GGUF: cstr/parakeet-tdt-0.6b-v3-GGUF
CrispASR supports 11 ASR backends in the same binary — Parakeet is one of the fastest for multilingual use with free word timestamps.