cstr/CrispASR

CrispASR Space

Gradio wrapper around the CrispASR HTTP server, packaged for Hugging Face Docker Spaces. One container runs the C++ inference engine on :8080 and the Gradio UI on :7860.

What's exposed

Models hot-swap through POST /load — only one model is resident at a time, so switching backends triggers a download (first use) and a load (every use). The first cold download for each backend lives in /cache.

The larger speech-LLM backends in CrispASR (Voxtral 2.5 GB, MiMo-ASR 4.5 GB, Granite-4.1 3 GB, omniasr-llm) are deliberately omitted from this demo — they exceed the free-tier (16 GB) RAM ceiling once Gradio + Python + KV cache overhead is accounted for. To run them, build the image locally.

Environment variables

• —CRISPASR_MODEL=/models/model.gguf (overridden by /load requests)
• —CRISPASR_BACKEND=whisper (initial backend; UI swaps later)
• —CRISPASR_LANGUAGE=auto (default language for transcription)
• —CRISPASR_AUTO_DOWNLOAD=1 (1 → resolve -m auto from the registry)
• —CRISPASR_CACHE_DIR=/cache (auto-download landing zone)
• —CRISPASR_SAMPLES_DIR=/space/samples (bundled jfk.wav etc.)
• —CRISPASR_API_KEYS= (optional comma-separated keys; protects every /v1/*)
• —CRISPASR_EXTRA_ARGS= (extra CLI flags forwarded verbatim, e.g. --vad --punc-model auto)

Local build / run

docker build -f hf-space/Dockerfile -t crispasr-hf-space .

docker run --rm -p 7860:7860 -p 8080:8080 \
  -e CRISPASR_BACKEND=whisper \
  -e CRISPASR_AUTO_DOWNLOAD=1 \
  crispasr-hf-space

Persist the model cache between runs:

docker volume create crispasr-cache
docker run --rm -p 7860:7860 -p 8080:8080 \
  -e CRISPASR_AUTO_DOWNLOAD=1 \
  -v crispasr-cache:/cache \
  crispasr-hf-space

Adjust build parallelism with --build-arg CRISPASR_BUILD_JOBS=8. The Dockerfile compiles two binaries from the CrispASR repo: crispasr (server + ASR/TTS) and crispasr-lid (text language ID).

Workflow notes

• —First transcription of a chosen backend triggers an HF download into /cache. Whisper-base (~147 MB) and Moonshine-tiny (~37 MB) feel instant; Cohere (~550 MB) and Qwen3-ASR (~500 MB) take a minute.
• —Swap order matters. Loading a TTS backend evicts the ASR backend and vice versa. Use the Transcribe tab's Load model button to swap back.
• —Voices: Kokoro's built-in voicepacks (af_heart, af_bella, am_michael, df_victoria, …) are baked into the auto-downloaded GGUF. Drop extra *.gguf / *.wav files into $VOICE_DIR if you want GET /v1/voices to list them; the Space doesn't ship a voice dir by default.
• —Long audio: enable VAD chunking with CRISPASR_EXTRA_ARGS=--vad so every backend processes minute-long files without truncation.
• —Word timestamps on LLM-style backends (qwen3, cohere) need an external CTC aligner — outside the scope of this free-tier demo, see docs/cli.md in the main repo.