Vosk Medium English (Batch)

Scoring

Primary metric: wer · Tie-breakers: real_time_factor, energy_wh

Metrics: wer, wer_substitutions, wer_deletions, wer_insertions, cer, cer_substitutions, cer_deletions, cer_insertions, real_time_factor, cpu_pct, ram_mb, energy_wh, temperature_c

Changelog

• v1.0.0 — Initial profile (larger per-language vosk model, alongside the existing small tier).
• v1.0.1 — Declare wer/cer substitution/deletion/insertion breakdown metrics, now that the runner computes them.
• v1.0.2 — Declare model.vad (false -- this adapter never applies it) and model.context_reset (per_utterance -- the adapter constructs a fresh KaldiRecognizer per utterance, no shared decode state to carry across calls in the first place) so a reader never has to infer these from adapter code.

Full definition

{
  "id": "vosk-medium-en-batch",
  "version": "1.0.2",
  "title": "Vosk Medium English (Batch)",
  "benchmark_type": "batch",
  "language": "en-US",
  "runtime": {
    "name": "vosk",
    "min_version": "0.3.44"
  },
  "model": {
    "name": "vosk-model-en-us-0.22",
    "vad": false,
    "context_reset": "per_utterance"
  },
  "configuration": {
    "threads": 4
  },
  "normalization": {
    "lowercase": true,
    "remove_punctuation": true,
    "expand_numbers": true,
    "ruleset_id": "goesb-en-v1"
  },
  "scoring": {
    "primary_metric": "wer",
    "tie_breakers": [
      "real_time_factor",
      "energy_wh"
    ]
  },
  "metrics": [
    "wer",
    "wer_substitutions",
    "wer_deletions",
    "wer_insertions",
    "cer",
    "cer_substitutions",
    "cer_deletions",
    "cer_insertions",
    "real_time_factor",
    "cpu_pct",
    "ram_mb",
    "energy_wh",
    "temperature_c"
  ],
  "changelog": [
    {
      "version": "1.0.0",
      "notes": "Initial profile (larger per-language vosk model, alongside the existing small tier)."
    },
    {
      "version": "1.0.1",
      "notes": "Declare wer/cer substitution/deletion/insertion breakdown metrics, now that the runner computes them."
    },
    {
      "version": "1.0.2",
      "notes": "Declare model.vad (false -- this adapter never applies it) and model.context_reset (per_utterance -- the adapter constructs a fresh KaldiRecognizer per utterance, no shared decode state to carry across calls in the first place) so a reader never has to infer these from adapter code."
    }
  ]
}