Blame for README.md

---
license: other
language:
- en
task_categories:
- automatic-speech-recognition
pretty_name: High-quality ASR soup (transcripts and metadata)
size_categories:
- 1M<n<10M
configs:
- config_name: default
  data_files:
  - split: train
    path: data/*/*.parquet
- config_name: common_voice
  data_files:
  - split: train
    path: data/common_voice/*.parquet
- config_name: spgispeech
  data_files:
  - split: train
    path: data/spgispeech/*.parquet
- config_name: gigaspeech_audiobook
  data_files:
  - split: train
    path: data/gigaspeech_audiobook/*.parquet
- config_name: ami
  data_files:
  - split: train
    path: data/ami/*.parquet
---

# highqualASRsoup

Transcripts and metadata of an English ASR training mix built from four public corpora. **No audio is included**: each
row carries the ids needed to fetch the audio from the upstream release.

| Subset | Upstream | Rows | Share of rows | Hours | Transcript |
|---|---|---:|---:|---:|---|
| `common_voice` | Common Voice English 26.0, train | 1,147,340 | 30.14 % | 1,812.4 | as released |
| `spgispeech` | SPGISpeech L, train | 1,926,805 | 50.62 % | 4,900.2 | as released |
| `gigaspeech_audiobook` | GigaSpeech L, audiobook-source rows | 623,565 | 16.38 % | 650.0 | rewritten (see below) |
| `ami` | AMI IHM, train | 108,502 | 2.85 % | 77.9 | rewritten (see below) |
| **Total** | | **3,806,212** | | **7,440.5** | |

## Transcript refinement

GigaSpeech and AMI ship upper-case transcripts; GigaSpeech also uses symbolic punctuation tokens (`<COMMA>`, `<PERIOD>`, ...).
For those two subsets every transcript was rewritten into normal written English (casing and punctuation) by
`gpt-5.6-luna` (prompt version `20260907-v2`); the upstream transcript is kept in `original_text`. Common Voice and
SPGISpeech transcripts are as released (`original_text` and `text_patch_model` are null), except 37 Common Voice rows
whose text had been merged with the following TSV record (tab/newline in the field) during the build and was re-read
literally from the release's `train.tsv`.

Example (`gigaspeech_audiobook`): `SIX TOMATOES <PERIOD>` → `Six tomatoes.`

## Columns

| Column | Meaning |
|---|---|
| `utt_id` | sha1 of `"<corpus>|<id>"`; stable row key |
| `corpus` | subset name |
| `id` | upstream utterance id: CV clip file name, SPGISpeech `wav_filename`, GigaSpeech `segment_id`, AMI `audio_id` |
| `text` | transcript used for training (refined for GigaSpeech / AMI) |
| `original_text`, `text_patch_model` | upstream transcript and rewrite model (GigaSpeech / AMI only) |
| `speaker` | CV `client_id`, AMI `speaker_id`, GigaSpeech `speaker` (upstream value, usually `N/A`) |
| `group` | leakage group for speaker/recording-disjoint splits: CV client, SPGISpeech call, GigaSpeech book (`audio_id`), AMI speaker |
| `duration_seconds` | from metadata: SPGISpeech WAV size (16 kHz, 16-bit), CV `clip_durations.tsv`, GigaSpeech / AMI `end_time - begin_time` |
| `begin_time`, `end_time` | segment times in the source recording (GigaSpeech, AMI) |
| `audio_id`, `meeting_id`, `microphone_id` | upstream recording ids (GigaSpeech book, AMI meeting / headset; CV clip) |
| `title`, `url`, `source`, `category`, `original_full_path` | GigaSpeech audiobook metadata, as released |
| `cv_source_dataset`, `wav_filesize` | CV release tag; SPGISpeech WAV size in bytes |
| `source_shard`, `source_row` | Parquet file and row in the soup build (SPGISpeech, GigaSpeech and AMI file names follow the upstream Hub shard names; Common Voice files are the build's own metadata shards) |

## Getting the audio

Join on `id` with the upstream release: Common Voice English 26.0 (`clips/<id>`), SPGISpeech
(`kensho/spgispeech`, config `L`, `wav_filename`), GigaSpeech (`speechcolab/gigaspeech`, config `l`, `segment_id`),
AMI (`edinburghcstr/ami`, config `ihm`, `audio_id`).

## License

Each subset remains under its upstream license and terms of use: Common Voice (CC0-1.0), AMI (CC BY 4.0), SPGISpeech and
GigaSpeech (gated on the Hub; accept their terms there). Use of these transcripts is subject to the same terms.