kdeps-components-transcribe

version: 1.0.1
requires:
    schemas/v1: ""
files:
    component.yaml: |
        apiVersion: kdeps.io/v1
        kind: Component
        metadata:
          name: transcribe
          description: Transcribe audio or video to text with a Whisper API or offline with whisper.cpp
          version: "1.0.1"
        interface:
          inputs:
            - name: file
              type: string
              required: true
              description: Path to the audio/video file. API backends take mp3, mp4, mpeg, mpga, m4a, wav, webm; whisper-cpp takes flac, mp3, ogg, wav.
            - name: backend
              type: string
              default: openai
              description: openai, groq, local (a self-hosted OpenAI-compatible Whisper server) or whisper-cpp (offline, no API key)
            - name: model
              type: string
              default: whisper-1
              description: API model (groq - whisper-large-v3); ignored by whisper-cpp
            - name: baseURL
              type: string
              default: ""
              description: Overrides the backend's API base URL (API backends only)
            - name: language
              type: string
              default: ""
              description: ISO-639-1 code, e.g. en; empty = auto-detect (whisper-cpp defaults to en)
            - name: prompt
              type: string
              default: ""
              description: Spelling hints for names and terms (API backends)
            - name: responseFormat
              type: string
              default: text
              description: text, json, verbose_json, srt or vtt (API backends)
            - name: temperature
              type: string
              default: ""
              description: Sampling temperature 0-1 (API backends)
            - name: modelPath
              type: string
              default: ""
              description: whisper-cpp GGML model; empty = ggml-base.en.bin, downloaded once to the models dir
        resources:
          # API backends: POST the file as multipart form data.
          - actionId: transcribeAPI
            name: Transcribe (Whisper API)
            validations:
              skip:
                - "get('transcribe.backend') == 'whisper-cpp'"
            httpClient:
              method: POST
              url: "{{ get('transcribe.baseURL') != '' ? get('transcribe.baseURL') : (get('transcribe.backend') == 'groq' ? 'https://api.groq.com/openai/v1' : (get('transcribe.backend') == 'local' ? 'http://localhost:8080/v1' : 'https://api.openai.com/v1')) }}/audio/transcriptions"
              headers:
                Authorization: "Bearer {{ get('transcribe.backend') == 'groq' ? env('GROQ_API_KEY') : env('OPENAI_API_KEY') }}"
              files:
                file: "{{ get('transcribe.file') }}"
              data:
                model: "{{ get('transcribe.model') }}"
                response_format: "{{ get('transcribe.responseFormat') }}"
                language: "{{ get('transcribe.language') }}"
                prompt: "{{ get('transcribe.prompt') }}"
                temperature: "{{ get('transcribe.temperature') }}"
              failOnError: true
              timeout: 300s
          # whisper-cpp: fetch the default model once, then run whisper-cli offline.
          - actionId: transcribeModel
            name: Download whisper.cpp model
            validations:
              skip:
                - "get('transcribe.backend') != 'whisper-cpp' || get('transcribe.modelPath') != '' || fileExists(env('KDEPS_MODELS_DIR', env('HOME') + '/.kdeps/models') + '/ggml-base.en.bin')"
            httpClient:
              method: GET
              url: https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin
              saveTo: "{{ env('KDEPS_MODELS_DIR', env('HOME') + '/.kdeps/models') }}/ggml-base.en.bin"
              failOnError: true
              timeout: 900s
          - actionId: transcribeOffline
            name: Transcribe (whisper.cpp)
            validations:
              skip:
                - "get('transcribe.backend') != 'whisper-cpp'"
            exec:
              command: whisper-cli
              args:
                - "-m"
                - "{{ get('transcribe.modelPath') != '' ? get('transcribe.modelPath') : env('KDEPS_MODELS_DIR', env('HOME') + '/.kdeps/models') + '/ggml-base.en.bin' }}"
                - "-f"
                - "{{ get('transcribe.file') }}"
                - "-l"
                - "{{ get('transcribe.language') != '' ? get('transcribe.language') : 'en' }}"
                - "-nt"
                - "-np"
              timeout: 600s