Transcribe

Transcribe speech to IPA using neurlang/ipa-whisper-small via the whisper.cpp inference server.

Usage

POST audio to /inference (multipart form field file):

curl -X POST http://127.0.0.1:8080/inference \
  -F "file=@input.wav" \
  -F "temperature=0.0" \
  -F "response_format=json"

Use 16 kHz mono WAV for best results. Convert with:

ffmpeg -i input.mp3 -ar 16000 -ac 1 -c:a pcm_s16le input.wav

Response is JSON; IPA text is in the segments (e.g. /ˌɪntərˈnæʃənəl/).

---

Setup

1. Model (one-time)

Convert and quantize the HuggingFace IPA model to ggml. See scripts/README.md for details.

./scripts/run-convert-ipa-to-ggml.sh              # → ggml (~466 MB)
export WHISPER_CPP_BUILD="$(pwd)/whisper.cpp/build"
./scripts/run-convert-ipa-to-ggml.sh --quantize   # → models/ggml-ipa-whisper-small-q5_0.bin (~182 MB)

Repo uses Git LFS for models/*.bin and bin/whisper-server. One-time: git lfs install. Then git add models/ggml-ipa-whisper-small-q5_0.bin and commit.

2. Build and run server

git submodule update --init
./scripts/build-server.sh
./scripts/run-server.sh

Server: http://0.0.0.0:8080. Install ffmpeg for non-WAV uploads (optional).

---

Deploy (Vercel)

Vercel’s builder has no cmake, so the whisper-server binary is prebuilt and committed via Git LFS (same approach as the model).

• Prebuild binary (one-time or when updating whisper.cpp):
  From repo root, run ./scripts/build-server-linux.sh (requires Docker). That produces a Linux x64 binary in bin/whisper-server. Then:

  git lfs install   # if not already
  git add bin/whisper-server
  git commit -m "Update prebuilt whisper-server (Linux)"
  git push

• Model: Tracked with Git LFS; Vercel runs git lfs pull in the install step.
• Build: npm run build uses the prebuilt bin/whisper-server from the repo (no cmake on Vercel). api/inference.ts proxies to it. Send WAV (16 kHz mono); no ffmpeg in bundle.
• Deploy: npm i -g vercel then vercel. Env: MODEL_PATH, BIN_DIR (optional).
• Limits: 250MB bundle, 1GB memory (in vercel.json). Use smaller model if needed.