README.md

metadata

title: Auris Nigerian Accented ASR
emoji: 🎙️
colorFrom: purple
colorTo: gray
sdk: gradio
app_file: app.py
pinned: false

Auris transcription service

Wraps NCAIR1/NigerianAccentedEnglish (a Whisper-small fine-tune on Nigerian-accented English) as a Gradio app running on free ZeroGPU hardware, for the Auris app's chunked recording pipeline.

API

Exposed as the /transcribe endpoint (see the Space's "Use via API" link once running).

• Inputs: audio (file/blob), secret (string — must match the API_SECRET below)
• Output: transcript text
• Errors (bad secret, no audio, transcription failure) surface as Gradio errors, which the @gradio/client caller sees as a status event with stage: "error".

Called from Node via @gradio/client's Client.submit(), not raw HTTP — this gives the caller real queue/cold-start visibility (stage: "pending", eta, position) instead of a single opaque request.

Configuration

Set both of these as Space secrets (Settings → Variables and secrets):

• HF_TOKEN — a HuggingFace access token (Read scope, with "Read access to contents of all public gated repos you can access" enabled) belonging to an account that has requested/been granted access on the model page. Without this the app starts but every call fails with a gated-repo error.
• API_SECRET — a shared secret of your choosing, so random visitors to this public Space can't burn through your ZeroGPU quota. The calling backend must send the same value as the secret input.

Hardware

Must be created with ZeroGPU hardware (the only free GPU tier as of writing). Model weights load once at startup (CPU); each call moves the model to the allocated GPU for the duration of that call and back to CPU afterward, since ZeroGPU only grants GPU access for the lifetime of a single @spaces.GPU-decorated call.

If the build or first call fails with a torch/CUDA compatibility error, check HF's current ZeroGPU docs for the currently-supported torch version range and pin it in requirements.txt — package versions drift over time and aren't pinned here.