Whisper Medium ES (Streaming)

Scoring

Primary metric: first_partial_latency · Tie-breakers: streaming_responsiveness, real_time_factor

Metrics: wer, wer_substitutions, wer_deletions, wer_insertions, real_time_factor, cpu_pct, ram_mb, first_partial_latency, first_final_latency, end_of_speech_latency, update_frequency, partial_stability, streaming_responsiveness

Changelog

• v1.0.0 — Initial profile (bulk-generated streaming set, matching batch's language/engine/size coverage).
• v1.0.1 — Declare wer/cer substitution/deletion/insertion breakdown metrics, now that the runner computes them.
• v1.0.2 — Declare model.context_reset (per_utterance) explicitly -- confirmed empirically (not assumed) that this adapter's Python API already has no cross-utterance context-bleed risk; documents that fact instead of leaving it implicit.

Full definition

{
  "id": "whisper-medium-es-streaming",
  "version": "1.0.2",
  "title": "Whisper Medium ES (Streaming)",
  "benchmark_type": "streaming",
  "language": "es-419",
  "runtime": {
    "name": "faster-whisper",
    "min_version": "1.0.0"
  },
  "model": {
    "name": "whisper-medium",
    "beam_size": 5,
    "temperature": 0,
    "quantization": "int8",
    "vad": true,
    "context_reset": "per_utterance"
  },
  "configuration": {
    "threads": 4,
    "chunk_ms": 1000
  },
  "normalization": {
    "lowercase": true,
    "remove_punctuation": true,
    "expand_numbers": true,
    "ruleset_id": "goesb-es-v1"
  },
  "scoring": {
    "primary_metric": "first_partial_latency",
    "tie_breakers": [
      "streaming_responsiveness",
      "real_time_factor"
    ]
  },
  "metrics": [
    "wer",
    "wer_substitutions",
    "wer_deletions",
    "wer_insertions",
    "real_time_factor",
    "cpu_pct",
    "ram_mb",
    "first_partial_latency",
    "first_final_latency",
    "end_of_speech_latency",
    "update_frequency",
    "partial_stability",
    "streaming_responsiveness"
  ],
  "overridable": {
    "beam_size": {
      "allowed": [
        1,
        2,
        4,
        5,
        8
      ]
    },
    "vad": {},
    "quantization": {
      "allowed": [
        "int8",
        "float32"
      ]
    },
    "threads": {
      "range": {
        "min": 1,
        "max": 16
      }
    },
    "chunk_ms": {
      "allowed": [
        250,
        500,
        1000,
        2000
      ]
    }
  },
  "changelog": [
    {
      "version": "1.0.0",
      "notes": "Initial profile (bulk-generated streaming set, matching batch's language/engine/size coverage)."
    },
    {
      "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.context_reset (per_utterance) explicitly -- confirmed empirically (not assumed) that this adapter's Python API already has no cross-utterance context-bleed risk; documents that fact instead of leaving it implicit."
    }
  ]
}