[asr]

Final transcription uses a local GGML Whisper model through whisper_rs.

[asr]
backend = "whisper_rs"
model_path = "~/.local/share/skald/models/ggml-large-v3-turbo-q5_0.bin"
language = "en"
threads = 8
gpu = true
gpu_backend = "cuda"

[asr.lifecycle]
mode = "keep_warm"
warm_on_daemon_start = true
idle_unload_seconds = 900

[asr.hallucination_filter]
enabled = true
phrases = [
  "thank you.",
  "thanks for watching.",
  "subtitles by*",
  "subtitle by*",
  "captioned by*",
]

[asr] options

Option	Type	Default	Description
backend	string	"whisper_rs"	ASR engine. Only whisper_rs is supported in v1.
model_path	string	large turbo q5 path under model_dir	Path to a GGML .bin model file. Tilde paths are expanded. File must exist before transcription (doctor warns if missing).
language	string	"en"	Whisper language code passed to the model.
threads	integer	8	CPU threads for inference (used for CPU paths and non-GPU work inside the backend).
gpu	boolean	true	Request GPU acceleration when skaldd was built with CUDA (just release-cuda). If true on a CPU-only build, model load fails with an unsupported-feature error.
gpu_backend	string	"cuda"	Reserved for future backend selection. Currently GPU use is controlled by gpu and the CUDA compile feature, not this field.

[asr.lifecycle] options

Option	Type	Default	Description
mode	string	"keep_warm"	keep_warm: keep model loaded between jobs (lower latency, higher memory). on_demand: load before each transcription and unload after (lower idle memory).
warm_on_daemon_start	boolean	true	When mode = "keep_warm", load the model when the daemon starts. Ignored for on_demand.
idle_unload_seconds	integer	900	When mode = "keep_warm" and the model is loaded, unload after this many seconds without use. Set 0 to disable idle unload.

[asr.hallucination_filter] options

Filters hallucination phrases Whisper sometimes emits on silence or noise. The filter applies only to transcripts of five words or fewer; longer output is never dropped by this check.

Each phrase is compared after normalization: internal whitespace is collapsed, case is folded to lowercase, and leading or trailing punctuation is stripped. A phrase ending in * is a prefix matcher (* is removed before normalization); all other phrases require an exact normalized match.

Option	Type	Default	Description
enabled	boolean	true	When true, drop short transcripts that match a listed phrase.
phrases	array of strings	see defaults	Phrases to treat as empty output. Suffix * for prefix match; otherwise exact after normalization.

Default phrases:

phrases = [
  "thank you.",
  "thanks for watching.",
  "subtitles by*",
  "subtitle by*",
  "captioned by*",
]

Model selection

• Use Setup wizard to benchmark candidates, or manage them directly with skald models.
• Use skald models recommend for a read-only recommendation based on CPU, RAM, NVIDIA/CUDA availability, installed catalog models, and current config.
• CPU-safe recommendation: small.en with gpu = false and on_demand.
• NVIDIA recommendation: large-v3-turbo-q5 for final transcription and small.en-q5 for text preview, after validating CUDA performance and VRAM use on the target system.
• Visualizer-only overlay mode does not require a preview model.
• After changing model_path or gpu, restart the daemon or run skald asr restart.

Catalog downloads are source-controlled metadata, not remote configuration. Files are streamed to the model filesystem, verified by exact size and SHA-256, then renamed into place. Paths outside the catalog remain supported but are reported as unverified and are never deleted by managed cleanup.

CPU-only selection:

skald models recommend
skald models install small.en
skald models select small.en
skald models select-preview small.en
skald config profile cpu-safe

NVIDIA/CUDA selection:

skald models recommend
skald models install large-v3-turbo-q5
skald models install small.en-q5
skald models select large-v3-turbo-q5
skald models select-preview small.en-q5
skald config profile power-user-nvidia

Notes

• Preview uses a separate small model configured in [preview], not asr.model_path.
• Benchmark timings: Linux benchmarks.