Add ASR release table v1 (6,448,108 rows, 43 languages) and dataset card

README.md

@@ -239,19 +239,21 @@ configs:
 ## Dataset summary
 
 This dataset contains 6,448,108 sentence-level speech clips cut from 46,373 YouTube videos
-in 43 languages, 29,751 hours in total. Each row describes one clip: its position in
-the source video (`video_id`, `start`, `end`), an automatic transcript produced with Whisper
-large-v3-turbo, a language label predicted from the transcript, and the attribution of the
-source video. The audio is not included.
+in 43 languages, 29,751 hours in total. The videos come from the YouTube collection
+of [Mumospee V2](https://huggingface.co/datasets/meetween/mumospee_V2) (`youtube_meeting` split): 33,275 of the 46,373 videos
+(71.8%) are in that split, and the others are videos of the same collection that are
+not included in it.
+
+Each row describes one clip: its position in the source video (`video_id`, `start`, `end`), an
+automatic transcript produced with Whisper large-v3-turbo, a language label predicted from the
+transcript, and the attribution of the source video. The audio is not included.
 
 All source videos were published under the Creative Commons Attribution license. The license
 of every video was checked with the YouTube Data API on 2026-10-15; videos that were no longer
 published under this license on that date are not included.
 
-The rows follow the shared metadata schema of the speechLMM-v2 data preparation pipeline (62
-columns), followed by 7 added columns. This is release v1: it contains all crawled videos in
-languages other than English and a sample of about 8,000 hours of English videos. The remaining
-English videos are planned for a later release.
+This is release v1: it contains the videos in languages other than English and a sample of about
+8,000 hours of English videos. The remaining English videos are planned for a later release.
 
 ## Languages and configurations
 
@@ -375,61 +377,38 @@ transcript is shortened here, and `attribution`, which is stored as a JSON strin
 
 ### Columns
 
-The first 62 columns are the speechLMM-v2 schema; selecting them gives rows in the exact
-speechLMM format. Only the following columns are filled; all other schema columns are null.
-
-| Column | Content |
-|---|---|
-| `sample_id` | `youtube_` + clip ID |
-| `task` | `["asr"]` |
-| `split` | `train` |
-| `source_dataset` | `youtube_clips_v1` |
-| `license` | `CC-BY-4.0` |
-| `audio_input_path` | null: the audio is not included |
-| `audio_input_duration` | Clip duration in seconds |
-| `audio_input_sample_rate`, `audio_input_num_channels`, `audio_input_codec` | 16000, 1, `flac`: the format of the clip audio from which the transcript was made |
-| `text_output` | Whisper transcript (cased and punctuated) |
-| `num_words_output` | Number of words in the transcript |
-| `language_text_output`, `language_audio_input` | Language of the transcript (ISO 639-1); null if it could not be identified. The label is predicted from the text; it is also given as `language_audio_input` because the speechLMM pipeline requires that column |
-| `lid_text_predicted`, `language_confidence_text` | Language label of the transcript (fastText `lid.176`; `und` if not identified) and its confidence |
-| `words_per_second`, `chars_per_second` | Speaking rate computed from the transcript and the duration |
-
-Added columns:
-
-| Column | Type | Content |
-|---|---|---|
-| `clip_id` | string | `<video_id>_<index>`; the index is the position of the clip in the video, from 0 |
-| `video_id` | string | YouTube video ID |
-| `start`, `end` | float64 | Start and end of the clip in the source video, in seconds |
-| `is_gold_transcript` | bool | `false` for all rows: the transcripts are model output, not human references |
-| `flags` | list of strings | Quality warnings: `fallback_cut` (the clip ends at a pause, not at the end of a sentence), `lang_und` (the language of the transcript could not be identified) |
-| `attribution` | string (JSON) | `title`, `uploader`, `channel_url` and `video_id` of the source video |
-
-## Dataset creation
-
-### Source data
-
-The source is a collection of public YouTube videos published under the Creative Commons
-Attribution license, crawled in 2026 together with their caption tracks. The captions were used
-only to select videos and estimate their length; they are not used as transcripts.
-
-### Processing
-
-1. **Transcription.** The audio of each video was converted to 16 kHz mono and transcribed with
-   faster-whisper (large-v3-turbo) with word timestamps and voice activity detection.
-2. **Segmentation.** The words were grouped into clips of 3 to 30 seconds (target 12 seconds)
-   that end at the end of a sentence. If no sentence ends before 30 seconds, the clip ends at the
-   longest pause (`fallback_cut`). A margin of 0.15 seconds is added at both ends, so a few clips
-   are slightly longer than 30 seconds.
-3. **Quality filter.** Videos with degenerate transcription output or with few sentence ends were
-   removed as a whole.
-4. **Language identification.** Each transcript was labelled with fastText `lid.176`; the video
-   language is the label with the largest share of speech in the video, weighted by the
-   confidence of the labels.
-5. **Deduplication.** Videos listed under more than one crawl language were kept once.
-6. **License check.** The license of every video was checked with the YouTube Data API on
-   2026-10-15. Of the 47,201 videos kept by the previous steps, 828 (98,886 clips) were
-   removed: 625 videos deleted, private or blocked; 203 videos now under the standard YouTube license.
+The rows use the dataset format of [speechLMM-v2](https://github.com/Meetween/speechlmm-v2), a shared format for speech
+and multimodal training data in which each row names its task and uses generic input, output,
+language and quality columns. The table lists the 25 columns used in this dataset; all
+other columns of the format are null.
+
+| Column | Type | Value in this dataset | Explanation |
+|---|---|---|---|
+| `sample_id` | string | `youtube_<clip_id>` | Unique ID of the row |
+| `task` | list of strings | `["asr"]` | Task of the row: automatic speech recognition |
+| `split` | string | `train` | Dataset split; all rows are training data |
+| `source_dataset` | string | `youtube_clips_v1` | Name of the source corpus, used when corpora are combined |
+| `license` | string | `CC-BY-4.0` | License of the row |
+| `audio_input_path` | string | null | Path of the input audio; null because the audio is not included |
+| `audio_input_duration` | float64 | seconds, e.g. `15.75` | Duration of the clip |
+| `audio_input_sample_rate` | int64 | `16000` | Sample rate in Hz of the clip audio from which the transcript was made |
+| `audio_input_num_channels` | int64 | `1` | Number of audio channels (mono) |
+| `audio_input_codec` | string | `flac` | Codec of the clip audio |
+| `text_output` | string | transcript | Target text: the Whisper large-v3-turbo transcript of the clip, cased and punctuated |
+| `num_words_output` | int64 | e.g. `37` | Number of words in `text_output` |
+| `language_audio_input` | string | ISO 639-1 code or null | Language of the input audio, as metadata of the row. Set to the label predicted from the transcript (`lid_text_predicted`). Null if no language was identified (28,212 rows, 0.4%), so that tools treat the language as unknown instead of as a language code |
+| `language_text_output` | string | ISO 639-1 code or null | Language of `text_output`, as metadata of the row; same value as `language_audio_input`, and null in the same rows |
+| `lid_text_predicted` | string | ISO 639-1 code or `und` | Output of the language classifier (fastText `lid.176`) on the transcript. Unlike the two columns above, it records the classifier result and is never null: `und` (ISO 639-2 code for "undetermined") means that the classifier ran but identified no language. These rows also have the flag `lang_und` |
+| `language_confidence_text` | float64 | 0 to 1 | Confidence of `lid_text_predicted` |
+| `words_per_second` | float64 | e.g. `2.35` | Speaking rate: `num_words_output` divided by `audio_input_duration` |
+| `chars_per_second` | float64 | e.g. `13.8` | Characters of `text_output` per second of audio |
+| `clip_id` | string | `<video_id>_<index>` | ID of the clip; the index is the position of the clip in its video, from 0. Rows of other tasks built on the same clip use the same `clip_id` |
+| `video_id` | string | YouTube video ID | Source video: `https://www.youtube.com/watch?v=<video_id>` |
+| `start` | float64 | seconds | Start of the clip in the source video |
+| `end` | float64 | seconds | End of the clip in the source video |
+| `is_gold_transcript` | bool | `false` | Whether the transcript is a human reference; `false` because the transcripts are model output |
+| `flags` | list of strings | empty, `fallback_cut`, `lang_und` | Quality warnings. `fallback_cut`: no sentence ended within 30 seconds, so the clip ends at the longest pause instead of a sentence end. `lang_und`: the language of the transcript could not be identified |
+| `attribution` | string (JSON) | `title`, `uploader`, `channel_url`, `video_id` | Title, channel name, channel URL and ID of the source video, for the attribution required by the CC BY license |
 
 ## Licensing and attribution
 
@@ -440,24 +419,3 @@ source video was published on YouTube under the Creative Commons Attribution lic
 
 Uploaders can change the license of a video or remove it after the check. The license status
 refers to the date of the check.
-
-## Removal requests
-
-If you are the author of a video in this dataset and want it removed, open a discussion in the
-Community tab of this repository with the video ID or URL. Removed videos are excluded from later
-versions of the dataset.
-
-## Limitations
-
-- **The transcripts are machine-generated.** They come from Whisper large-v3-turbo and have not
-  been checked against human transcripts; their word error rate is not known. A model trained on
-  this data learns to reproduce Whisper output. Evaluate on test sets with human transcripts.
-- **Language labels are predicted from the text.** For about half of the videos (the first
-  processing batch), Whisper was set to decode in the language label of the crawl. If that label was wrong, the transcript
-  may not match the spoken language, and a text-based label cannot detect this.
-- **The language distribution is uneven.** The nine largest languages hold almost all hours;
-  many languages have less than 10 hours.
-- **Clip length.** About 2.4% of the clips are shorter than 3 seconds, mostly the last
-  clip of a video or clips cut at a pause.
-- **No speaker information and no evaluation split.** All rows are in `train`. Clips of the same
-  video share speakers and recording conditions; split by `video_id`, not by clip.