README.md

metadata

language:
  - ja
task_categories:
  - automatic-speech-recognition
license: other
pretty_name: JP-HomophoneBench
tags:
  - japanese
  - asr
  - homophone
  - phoneme
  - contextual-asr
  - context-biasing
  - speech-recognition
configs:
  - config_name: homophone8
    data_files:
      - split: train
        path: homophone8/train-00000-of-00001.parquet
      - split: validation
        path: homophone8/validation-00000-of-00001.parquet
      - split: test
        path: homophone8/test-00000-of-00001.parquet
  - config_name: homophone8-audio
    data_files:
      - split: test
        path: homophone8-audio/test-*
  - config_name: homophone8-research
    data_files:
      - split: train
        path: homophone8-research/train-00000-of-00001.parquet
      - split: validation
        path: homophone8-research/validation-00000-of-00001.parquet
      - split: test
        path: homophone8-research/test-00000-of-00001.parquet
dataset_info:
  config_name: homophone8-audio
  features:
    - name: audio_filepath
      dtype: 'null'
    - name: audio_ref
      struct:
        - name: config
          dtype: string
        - name: kind
          dtype: string
        - name: repo_id
          dtype: string
        - name: row_id
          dtype: string
        - name: split
          dtype: string
    - name: candidates
      list: json
    - name: category
      dtype: string
    - name: difficulty
      struct:
        - name: acoustic
          dtype: float64
        - name: context
          dtype: float64
        - name: lexical
          dtype: float64
        - name: phone_distance
          dtype: float64
    - name: group_id
      dtype: string
    - name: id
      dtype: string
    - name: metadata
      dtype: json
    - name: source
      struct:
        - name: audio_redistributed
          dtype: bool
        - name: config
          dtype: string
        - name: dataset
          dtype: string
        - name: license
          dtype: string
        - name: revision
          dtype: string
        - name: source_id
          dtype: string
        - name: split
          dtype: string
        - name: synthetic
          dtype: bool
    - name: split
      dtype: string
    - name: target
      dtype: json
    - name: text
      dtype: string
    - name: benchmark_split
      dtype: string
    - name: audio
      dtype:
        audio:
          sampling_rate: 16000
    - name: audio_origin
      dtype: string
    - name: audio_engine
      dtype: string
    - name: audio_engine_version
      dtype: string
    - name: audio_voice
      dtype: string
    - name: audio_voice_license
      dtype: string
    - name: audio_source_config
      dtype: string
    - name: audio_source_revision
      dtype: string
    - name: audio_sampling_rate
      dtype: int64
  splits:
    - name: test
      num_bytes: 3761077
      num_examples: 71
  download_size: 3748747
  dataset_size: 3761077

JP-HomophoneBench

A deterministic Japanese ASR benchmark index for separating eight error/disambiguation classes:

• exact_homophone
• near_homophone
• voicing
• long_vowel
• geminate
• moraic_nasal
• pitch_accent
• semantic_only

Important design rule

This repository is metadata-first. Source audio is not redistributed by default. Each row stores source repository/config/split/row identifiers so audio can be rehydrated under the original source license.

exact_homophone and semantic_only are intentionally separate. A phoneme scorer is not expected to solve semantic_only by itself.

HF repository: saeeew/JP-HomophoneBench Most recently published configuration: homophone8 Publication license policy: permissive

Available configurations

• homophone8
• homophone8-research

Published rows in this run

Splits

• train: 111
• validation: 9
• test: 11

Categories

• exact_homophone: 10
• geminate: 3
• long_vowel: 3
• moraic_nasal: 3
• near_homophone: 7
• pitch_accent: 92
• semantic_only: 10
• voicing: 3

Upstream sources

• HaitongSUN/prosodic-abx: 92
• JP-HomophoneBench/permissive-phonetic-core: 19
• manual-cc0-example: 20

Licensing and provenance

The combined benchmark uses license: other because rows retain their upstream license. Inspect source.license before redistribution or commercial use.

Loading

from datasets import load_dataset
ds = load_dataset("saeeew/JP-HomophoneBench", "homophone8")
print(ds)

Build summary

{
  "augmentation": "permissive-core8-v0.1",
  "categories": {
    "exact_homophone": 178,
    "geminate": 9,
    "long_vowel": 18,
    "moraic_nasal": 5,
    "near_homophone": 210,
    "pitch_accent": 92,
    "semantic_only": 178,
    "voicing": 17
  },
  "core8_missing": [],
  "licenses": {
    "cc-by-4.0": 92,
    "cc-by-nc-4.0": 576,
    "cc0-1.0": 39
  },
  "records": 707,
  "sha256": {
    "all.jsonl": "897940216f5f1130b0954e3381f0ec37d998575791f2b170e3add3c5b664eeb3",
    "test.jsonl": "a657d1accff04dd7c1c629a4ef1fe8232b11ed2425a962187feee5fff6aebb43",
    "train.jsonl": "93a3fa069fefada5564a2f550a9ce98579fe611da4c6afcb43b5e8e755f5f18d",
    "validation.jsonl": "12e5d99945e844d8be53aea4f2864a912b56bb907458291ed698d25cd60f564c"
  },
  "sources": {
    "HaitongSUN/prosodic-abx": 92,
    "IDEMITSU/hoiku-yougo-stt-ja": 576,
    "JP-HomophoneBench/permissive-phonetic-core": 19,
    "manual-cc0-example": 20
  },
  "splits": {
    "test": 161,
    "train": 475,
    "validation": 71
  }
}