scripted-speech

Scripted Speech (SCS)

Scripted Speech is the classic Common Voice dataset. Contributors read pre-written sentences aloud, and the community validates the recordings. New datasets are released approximately every quarter.

All voice contributions are released as part of datasets, regardless of validation status. From v25.0 on, clips that fail quality checks (over-length, corrupted, or missing audio files) are excluded during bundling; a per-locale problem clip report is included with each release for transparency. The clips are currently bundled using the embedded bundler in the public repo Common Voice - Bundler.

Release History

See the full Changelog for detailed release notes and new languages per release.

Total and Validated Hours

---
config:
    xyChart:
        width: 900
        height: 400
---
xychart-beta
    title "Scripted Speech: Total & Validated Hours"
    x-axis ["1","2","3","4","5.1","6.1","7","8","9","10","11","12","13","14","15","16.1","17","18","19","20","21","22","23","24","25"]
    y-axis "Hours" 0 --> 42000
    bar [1368,2366,2454,4257,7226,9283,13905,18243,20217,20817,24231,26119,27141,28117,28750,30328,31175,32121,32584,33154,33534,33815,35921,38932,41792]
    bar [1096,1872,1979,3401,5671,7335,11192,14122,14973,15234,16429,17127,17689,18651,19159,19915,20408,20943,21593,22106,22344,22640,24600,25886,28377]

Loading

Contributors

---
config:
    xyChart:
        width: 900
        height: 400
---
xychart-beta
    title "Scripted Speech: Total Contributors"
    x-axis ["1","2","3","4","5.1","6.1","7","8","9","10","11","12","13","14","15","16.1","17","18","19","20","21","22","23","24","25"]
    y-axis "Users" 0 --> 500000
    bar [42109,56059,57420,95798,138225,151434,191622,207602,252576,263879,271817,281069,288617,298724,302232,319703,330323,335780,338378,345996,350098,356074,361614,371058,375673]

Loading

Counts are summed per language — contributors active in multiple languages are counted once per language.

Language Count

---
config:
    xyChart:
        width: 900
        height: 400
---
xychart-beta
    title "Scripted Speech: Languages per Release"
    x-axis ["1","2","3","4","5.1","6.1","7","8","9","10","11","12","13","14","15","16.1","17","18","19","20","21","22","23","24","25"]
    y-axis "Languages" 0 --> 310
    line [19,28,29,40,54,60,76,87,93,96,100,104,108,112,114,120,124,129,131,133,134,137,286,289,290]

Loading

Release Summary

Release	Date	Languages	Total Hours	Validated Hours
v1	2019-02-25	19	1,368	1,096
v2	2019-06-11	28	2,366	1,872
v3	2019-06-24	29	2,454	1,979
v4	2019-12-10	40	4,257	3,401
v5.1	2020-06-22	54	7,226	5,671
v6.1	2020-12-11	60	9,283	7,335
v7.0	2021-07-21	76	13,905	11,192
v8.0	2022-01-19	87	18,243	14,122
v9.0	2022-04-27	93	20,217	14,973
v10.0	2022-07-04	96	20,817	15,234
v11.0	2022-09-21	100	24,231	16,429
v12.0	2022-12-07	104	26,119	17,127
v13.0	2023-03-09	108	27,141	17,689
v14.0	2023-06-23	112	28,117	18,651
v15.0	2023-09-08	114	28,750	19,159
v16.1	2023-12-06	120	30,328	19,915
v17.0	2024-03-15	124	31,175	20,408
v18.0	2024-06-14	129	32,121	20,943
v19.0	2024-09-13	131	32,584	21,593
v20.0	2024-12-06	133	33,154	22,106
v21.0	2025-03-14	134	33,534	22,344
v22.0	2025-06-20	137	33,815	22,640
v23.0	2025-09-05	286	35,921	24,600
v24.0	2025-12-05	289	38,932	25,886
v25.0	2026-03-09	290	41,792	28,377

About the Statistics

Statistics for each release are stored as JSON files in this directory. The JSON structure may have changed slightly from release to release, so if you plan on doing any comparisons you may need to normalize them between versions.

Any demographic split (i.e. gender, age, accent) is applied to the entire dataset, not just the validated set. Unless otherwise indicated, durations are measured in milliseconds, and file sizes are measured in bytes.

Archive Structure

Each downloaded .tar.gz file has the following structure, where {lang} represents the BCP 47 locale code for that language:

cv-corpus-{version}-{YYYY-MM-DD}-{lang}.tar.gz/
  cv-corpus-{version}-{YYYY-MM-DD}/
  └── {lang}/
      ├── README.md                (datasheet, since Corpus 25.0)
      ├── clips/
      │   └── *.mp3
      ├── dev.tsv
      ├── invalidated.tsv
      ├── other.tsv
      ├── test.tsv
      ├── train.tsv
      ├── validated.tsv
      ├── reported.tsv
      ├── clip_durations.tsv
      ├── validated_sentences.tsv
      └── unvalidated_sentences.tsv

TSV Fields

Each row of a clip TSV file (validated.tsv, invalidated.tsv, other.tsv, train.tsv, dev.tsv, test.tsv) represents a single audio clip:

• client_id -- hashed UUID of a given user
• path -- relative path of the audio file
• sentence -- transcription of the audio to be read aloud by the contributor
• sentence_id -- unique identifier for the sentence (since Corpus 17.0)
• sentence_domain -- domain classification(s) of the sentence (since Corpus 17.0)
• up_votes -- number of people who said audio matches the sentence
• down_votes -- number of people who said audio does not match the sentence
• age -- age bracket of the speaker - if provided*
• gender -- gender of the speaker - if provided*
• accents -- accent(s) of the speaker - if provided* (previously named accent but renamed to reflect multiple selections, since Corpus 17.0)
• variant -- language variant - if provided (since Corpus 13.0)
• locale -- locale code of the language (since Corpus 5.0)
• segment -- custom dataset segment, if applicable (since Corpus 5.0)

The train.tsv, dev.tsv, and test.tsv splits are produced by CorporaCreator and contain the same columns as validated.tsv.

*For a full list of age, gender, and accent options, see the demographics spec. These are only reported if the speaker opted in.

Additional TSV Files

clip_durations.tsv (since Corpus 16.1) -- clip filename and duration:

• clip -- clip filename
• duration[ms] -- duration of the clip in milliseconds

validated_sentences.tsv (since Corpus 17.0) -- sentences that have reached the validated threshold (two or more up votes):

• sentence_id -- unique identifier for the sentence
• sentence -- the sentence itself
• variant -- language variant token for the sentence, if provided (since Corpus 25.0)
• sentence_domain -- domain classification(s) of the sentence, if provided
• source -- origin of the sentence (user provided or from old files under server/data)
• is_used -- whether the sentence is still eligible for recording (sentences may be retired if they are incorrect, outdated, too similar to other sentences, or for other reasons via database migrations)
• clips_count -- number of clips recorded for this sentence

unvalidated_sentences.tsv (since Corpus 17.0) -- sentences that have not reached the validated threshold or have been rejected:

• sentence_id -- unique identifier for the sentence
• sentence -- the sentence itself
• variant -- language variant token for the sentence, if provided (since Corpus 25.0)
• sentence_domain -- domain classification(s) of the sentence, if provided
• source -- origin of the sentence (user provided or from old files under server/data)
• up_votes -- number of approving votes (since Corpus 25.0)
• down_votes -- number of rejecting votes (since Corpus 25.0)
• status -- pending (not yet decided) or rejected (2+ down votes exceeding up votes) (since Corpus 25.0)

Validation Categories

• validated -- clips with two or more validations where up_votes > down_votes
• invalidated -- clips with two or more validations where down_votes > up_votes, or three or more where down_votes = up_votes
• other -- clips without sufficient validations to determine their status

reported.tsv (since Corpus 5.0) -- sentences flagged by contributors:

• sentence -- text of the reported sentence
• sentence_id -- unique identifier for the sentence
• locale -- locale code
• reason -- report reason: offensive-language, grammar-or-spelling, different-language, difficult-pronounce

Note: reporting a sentence does not remove it from circulation. Reported sentences remain available for recording and validation. The reported.tsv file is provided for post-processing by dataset consumers.

Use for Machine Learning

We use the Corpora Creator tool to parse through metadata to generate train, dev, and test sets. The Corpora Creator eliminates duplication in clips and maximizes for speaker diversity.

Each train/dev/test set is generated non-deterministically, meaning they will vary from release to release even for minor updates. This is to avoid reproducing and perpetuating any demographic skews in each subsequent set. Note that total clips in these sets will most probably not add up to the total validated clips because of this limitation. Please check the repo to include multiple recordings per sentence (the -s flag) if you want to get as close as possible to the total validated clips.