Transcribe audio — vosk (option, not installed)

Transcribe an audio file with vosk (Kaldi-based), offline, from Python, with small per-language models. A declared OPTION: not installed in this repository.

Skills it satisfies

A skill states the capability; this Tool is one way to exercise it. Each skill’s declared input and output contract follows, where it has one.

library-ingestion

The BASIC ingestion flow every asset takes: an upload is accepted, its metadata goes into the knowledge graph, and the asset lands in library/<slug>/ if it is materialized. The two entry points (a drop in uploads/, or materializing an asset a remote graph lists), what happens to the upload afterwards, and why content-type methods refine this flow from above rather than living in it. Command: bun run cat ingest.

Source: cat-harness/library/skills/library-core/library-ingestion.md.

Also satisfied by: ingest-extended, ingest-stdlib, l1-complete-check, narrative-queue, openapi-ingest-rendered, openapi-ingest-source, transcribe-faster-whisper, transcribe-whisper-cpp.

The skill declares no I/O contract (schemas/skills/library-ingestion/{input,output}.schema.json).

Inputs and outputs

Each port names a shared type by IRI, so two Tools for one skill can be checked for interchangeability.

porttyperequireddescriptionconstraint
infileRepoPathyesstring · pattern ^[A-Za-z0-9._][A-Za-z0-9._/-]*$ · A path relative to the repository root. No .. segments.
inlanguageLocalestring · pattern ^[a-z]{2,3}(-[A-Za-z0-9]{2,8})*$ · A BCP-47 language tag, e.g. en or fr-CA
outtranscriptRepoPathWritten where -o names.string · pattern ^[A-Za-z0-9._][A-Za-z0-9._/-]*$ · A path relative to the repository root. No .. segments.

Processes and tasks

Steps this Tool can carry out, because each names a skill it satisfies:

processsteplanevia skill
Content acquisitionRoute it, and watch the queueAgentlibrary-ingestion
Basic ingestion — an upload to an asset catalogued in library/[Place the asset in library//](../../../processes/document-ingestion.html#step-Task_PlaceInLibrary)Ingestion Engine (agent, runs unattended)library-ingestion
Basic ingestion — an upload to an asset catalogued in library/Record it in library/ as referencedIngestion Engine (agent, runs unattended)library-ingestion
L1 document ingestion — a document to the L1 source knowledge graphBasic ingestion (harness)Ingestion Engine (agent, runs unattended)library-ingestion
DTH authoring1 · Ingest sources into library/Librarianlibrary-ingestion

How it is deployed and reached

  • Command line. From the repository root: vosk-transcriber.
  • GitHub Actions. No workflow in this checkout runs its command.

As an API: its operation in the OpenAPI document and the MCP tool list of its layer.

installcli: pip install vosk, plus a per-language model (small ≈ 40–50 MB)
runtimepython3, vosk
networkno
when to choose it
whenWhen size and speed matter more than accuracy: the lightest offline option, with a separate small model per language.
limitsNoticeably lower accuracy than the Whisper family on varied recordings; one model per language to fetch and keep.
costNot installed. Small wheel; ≈ 40–50 MB per language model.

Alternatives (same signature, a shared skill): transcribe-faster-whisper, transcribe-whisper-cpp.

What it is made of

Linked data

The Tool is published as its own JSON-LD document at its IRI, https://litlfred.github.io/folio-assistant/cat-harness/tools/transcribe-vosk.jsonld, partOf its layer’s document. The node below is the definition this page was drawn from, and is embedded in the page as application/ld+json.

{
  "@context": "https://litlfred.github.io/folio-assistant/cat-harness/cat-harness.context.jsonld",
  "@id": "https://litlfred.github.io/folio-assistant/cat-harness/tools/transcribe-vosk.jsonld",
  "@type": "https://litlfred.github.io/cat-harness/v0/ns/Tool",
  "partOf": "https://litlfred.github.io/folio-assistant/cat-harness/cat-harness.jsonld",
  "name": "transcribe-vosk",
  "title": "Transcribe audio — vosk (option, not installed)",
  "description": "Transcribe an audio file with vosk (Kaldi-based), offline, from Python, with small per-language models. A declared OPTION: not installed in this repository.",
  "install": {
    "cli": "pip install vosk, plus a per-language model (small ≈ 40–50 MB)"
  },
  "invoke": {
    "shell": "vosk-transcriber"
  },
  "io": {
    "inputs": [
      {
        "name": "file",
        "schema": "https://litlfred.github.io/folio-assistant/cat-harness/tool-types.schema.json#/$defs/RepoPath",
        "required": true,
        "arg": {
          "flag": "-i"
        }
      },
      {
        "name": "language",
        "schema": "https://litlfred.github.io/folio-assistant/cat-harness/tool-types.schema.json#/$defs/Locale",
        "required": false,
        "arg": {
          "flag": "-l"
        }
      }
    ],
    "outputs": [
      {
        "name": "transcript",
        "schema": "https://litlfred.github.io/folio-assistant/cat-harness/tool-types.schema.json#/$defs/RepoPath",
        "description": "Written where `-o` names."
      }
    ]
  },
  "satisfies": [
    "library-ingestion"
  ],
  "requirements": {
    "runtime": [
      "python3",
      "vosk"
    ],
    "network": false
  },
  "selection": {
    "when": "When size and speed matter more than accuracy: the lightest offline option, with a separate small model per language.",
    "limits": "Noticeably lower accuracy than the Whisper family on varied recordings; one model per language to fetch and keep.",
    "cost": "Not installed. Small wheel; ≈ 40–50 MB per language model."
  }
}

folio-assistant — code under the Apache License 2.0, documentation under CC BY 3.0. Generated schema and API references are rebuilt from source on every push to main.

deployed dev