Loading datasets...
Task: ASR
Release Date: 3/22/2026
Format: MP3
Size: 2.78 GB
Share
A collection of read speech recordings in Turkish (Türkçe).
Restrictions/Special Constraints
None provided.
Forbidden Usage
It is forbidden to attempt to determine the identity of speakers in the Common Voice datasets. It is forbidden to re-host or re-share this dataset.
Intended Use
This dataset is intended to be used for training and evaluating automatic speech recognition (ASR) models. It may also be used for applications relating to computer-aided language learning (CALL) and language or heritage revitalisation.
tr)This datasheet is for cv-corpus-25.0-2026-03-09 of the Mozilla Common Voice Scripted Speech dataset for Turkish [Türkçe - tr]. The dataset contains 126510 clips representing 135.29 hours of recorded speech (129.22 hours validated) from 1816 speakers, recorded from a text corpus of 413,900 sentences.
Turkish is the most widely spoken language among Turkic languages and has around 100 million L1 speakers, which makes it the 18th most spoken language. It is the national language of Turkey and one of two official languages of Cyprus, and secondary languages of some neighboring countries. Many smaller groups in other countries exist, through migrations or communities from Ottoman era. These smaller groups should usually be categorized as a variant.
There are currently no variants defined for Common Voice Turkish dataset. It is worth noting that, until now, this dataset focused on literary Turkish, often called "Turkish of Turkey". There are also some L2 voices, mostly from immigrants coming into the country, but these can be categorized as "foreign accents".
| Code | Accent | Clips | Speakers |
|---|---|---|---|
| - | 49,235 (38.9%) | 159 (8.8%) |
The dataset includes the following self-declared age and gender distributions. A coverage summary is shown below each table.
Self-declared gender information. The table shows clip and speaker counts with percentages. Speakers who did not declare a gender are listed as Unspecified. A dash (-) indicates zero.
| Code | Gender | Clips | Speakers |
|---|---|---|---|
| male_masculine | Male, masculine | 53,360 (42.2%) | 460 (25.3%) |
| female_feminine | Female, feminine | 40,592 (32.1%) | 102 (5.6%) |
| transgender | Transgender | - | - |
| non-binary | Non-binary | - | - |
| do_not_wish_to_say | Prefer not to say | - | - |
| - | Unspecified | 32,558 (25.7%) | 1,368 (75.3%) |
Gender declared: 93,952 of 126,510 clips (74.3%), 448 of 1,816 speakers (24.7%)
Self-declared age information. The table shows clip and speaker counts with percentages. Speakers who did not declare an age are listed as Unspecified. A dash (-) indicates zero.
| Code | Age | Clips | Speakers |
|---|---|---|---|
| teens | Teens | 1,957 (1.5%) | 55 (3.0%) |
| twenties | Twenties | 28,575 (22.6%) | 299 (16.5%) |
| thirties | Thirties | 10,839 (8.6%) | 163 (9.0%) |
| fourties | Fourties | 3,214 (2.5%) | 59 (3.2%) |
| fifties | Fifties | 8,940 (7.1%) | 30 (1.7%) |
| sixties | Sixties | 25,304 (20.0%) | 12 (0.7%) |
| seventies | Seventies | 4,187 (3.3%) | 2 (0.1%) |
| eighties | Eighties | 12,300 (9.7%) | 2 (0.1%) |
| nineties | Nineties | - | - |
| - | Unspecified | 31,194 (24.7%) | 1,325 (73.0%) |
Age declared: 95,316 of 126,510 clips (75.3%), 491 of 1,816 speakers (27.0%)
Clip buckets
| Bucket | Clips |
|---|---|
| Validated | 120,832 (95.5%) |
| Invalidated | 5,018 (4.0%) |
| Other | 660 (0.5%) |
Training splits
| Split | Clips |
|---|---|
| Train | 40,815 (33.8%) |
| Dev | 11,797 (9.8%) |
| Test | 11,819 (9.8%) |
Training split coverage: 64,431 of 120,832 validated clips (53.3%)
The dataset contains 120832 validated, 5018 invalidated, and 660 unresolved clips. The average clip duration is 3.85 seconds.
Actual values of text corpus from the previous version (please check the text corpus tab) show that:
Average Characters/Sentence of the whole corpus is 62.25, but validated set average is 35.915. Median values are 61 and 27 respectively.
Average Words/Sentence of the whole corpus is 8.379, but validated set average is 5.096. Median values are 8 and 4 respectively.
Shorter sentences in the older text corpus resulted in shorter recording durations, which is 3.8 seconds on the average (in the previous dataset).
This value is increasing with each version, and we aim to reach 5 sec, which is the ideal minimum of most SotA model architectures.
Validated sentences: 410,410
| Category | Count |
|---|---|
| Unvalidated sentences | 3,490 |
| Pending sentences | 40 |
| Rejected sentences | 3,450 |
| Reported sentences | 486 |
The corpus contains 413,900 sentences: 410,410 validated and 3,490 unvalidated (40 pending review, 3,450 rejected), with 486 reported for review.
Turkish uses an extended Latin alphabet.
Official Alphabet:
Lowercase: a b c ç d e f g ğ h ı i j k l m n o ö p r s ş t u ü v y z
Uppercase: A B C Ç D E F G Ğ H I İ J K L M N O Ö P R S Ş T U Ü V Y Z
Auxilary Characters (Arabic/Farsi loanwords): â î û Â Î Û
There follows a randomly selected sample of five sentences from the corpus.
Bitinya'da Olimpos Dağı'nda vaftiz edildiği manastırda keşiş oldu.
Seçilmesinden beş gün sonra ani bir şekilde öldü.
Tahta resim kalıbı üstüne çizdiği Fransa haritası türünün ilkidir.
Lafın gelişi.
Mesleğe başladığı sırada birinci Dünya Savaşı başladı ve silah altına alındı.
The whole history of the text-corpora added throughout the years can be found in this forum post (Turkish). They mainly consist of:
Pre 2021: SETimes => ~5k sentences (only these are recorded multiple times between 2018-2021, ~30h audio)
2021/10: Turkish proverbs => ~2.5k
2021-2022: Extracted sentences from books of Sabahattin Ali => ~18-19k sentences
2021-2024: Community generated conversational sentences => ~33k sentences
2023/11: Wikipedia random selection through cv-sentence-extractor => 348.5k
Please note that:
Not all of these sentences are recorded. The validated set only includes ~63k unique sentences.
Until inclusion of Wikipedia sentences, we did not put a minimum limit to the sentence length. Especially after adding short conversational sentences, the average recording duration dropped. For this reason, although they are descriptive statements, we put a 3 word/20 char minimum limit to Wikipedia sentences.
We currently have ~110k sentences extracted from public works of deceased authors, and 20k community generated sentences, waiting to be cross checked.
| Source | Sentences |
|---|---|
| https://tr.wikipedia.org | 348,494 (85.0%) |
| sentence-collector | 47,746 (11.6%) |
| setimes | 5,077 (1.2%) |
| Other | 8,833 (2.2%) |
Until this version, we did not work on specialized sentence domains, except entries coming from individuals and ~300 numbers we added.
| Code | Domain | Clips | Speakers |
|---|---|---|---|
| general | General | 2 (0.0%) | 2 (0.1%) |
| agriculture_food | Agriculture and Food | - | - |
| automotive_transport | Automotive and Transport | 1 (0.0%) | 1 (0.1%) |
| finance | Finance | - | - |
| service_retail | Service and Retail | - | - |
| healthcare | Healthcare | - | - |
| history_law_government | History, Law and Government | - | - |
| media_entertainment | Media and Entertainment | 1 (0.0%) | 1 (0.1%) |
| nature_environment | Nature and Environment | - | - |
| news_current_affairs | News and Current Affairs | - | - |
| technology_robotics | Technology and Robotics | - | - |
| language_fundamentals | Language Fundamentals | 12 (0.0%) | 4 (0.2%) |
Except entries coming from individuals, we followed the following procedure for bulk additions to the text-corpus:
Extraction using Common Voice rules (e.g. 14 words)
Using Google Sheets to create conversational sentences in volunteer groups
Every entry has been check twice by the datasheet author, then at least one person checked it again, fully.
Wikipedia data extracted using the cv-sentence-extractor has been done through a long and carefull process to keep the corpus quality high. We don't expect more than 2.5% error rate in these sentences, and most will be based on the bad grammer in originals, which we could not correct at that time due to the random nature of the algorithm.
Please note that Turkish in Common Voice does not have a special language based validator, it uses the defaults.
Because there has been no sentence validation, there are some non-alphabet characters in the text corpus, especially coming from the SETimes corpus via proper names. You may like to remove these.
Don't normalize extra characters used for Arabic/Farsi loanwords (â, î, û), normalizing them to a/ı/u will change the meaning and intonation.
There are some proper names containing x and w, there are few and can be removed. We usually keep them because they are also used in minority languages of Turkey and in some proper names.
Each row of a tsv file represents a single audio clip, and contains the following information:
client_id - hashed UUID of a given user
path - relative path of the audio file
text - supposed transcription of the audio
up_votes - number of people who said audio matches the text
down_votes - number of people who said audio does not match text
age - age of the speaker1
gender - gender of the speaker1
accents - accents of the speaker1
variant - variant of the language1
segment - if sentence belongs to a custom dataset segment, it will be listed here
prompt_upvotes - number of upvotes the sentence prompt received
prompt_reports - number of reports the sentence prompt received
is_edited - whether the clip's transcription has been edited
validated_sentences.tsvThe validated_sentences.tsv file contains one row per validated sentence in the text corpus:
sentence_id - unique identifier for the sentence
sentence - the sentence text
variant - the variant of the language
sentence_domain - the domain(s) the sentence belongs to
source - the source the sentence was collected from
is_used - whether the sentence is still in circulation for recording
clips_count - number of clips recorded for this sentence
unvalidated_sentences.tsvThe unvalidated_sentences.tsv file contains one row per unvalidated sentence in the text corpus:
sentence_id - unique identifier for the sentence
sentence - the sentence text
variant - the variant of the language
sentence_domain - the domain(s) the sentence belongs to
source - the source the sentence was collected from
up_votes - number of upvotes the sentence received
down_votes - number of downvotes the sentence received
status - current status of the sentence (pending or rejected)
Main Channels:
Social media channels used during campaigns:
Most info can be found in Turkish language in the Discourse Turkish sub-forum. Other discussions are in Discourse main forum in English. Current discussions are on the Telegram group.
"How to contribute?", "What to avoid?", and similar topics for the newcomers can be found in the following forum post in Turkish: The process, rights & wrongs and dataset improvement
If you want to contribute, please first join the Telegram group.
Our future plans include:
Adding more conversational sentences, validating extracted 110k sentences, adding longer sentences.
Providing domain based text-corpora
Adding pre-defined variants and accents
Adding validators
Prepare a global campaign
You can find more information about how to participate in the Common Voice Project on the following page:
Community Participation Guidelines
Bülent Özden
This dataset is released under the Creative Commons Zero (CC-0) licence. By downloading this data you agree to not determine the identity of speakers in the dataset.
For a full list of age, gender, and accent options, see the demographics spec. These will only be reported if the speaker opted in to provide that information. ↩ ↩2 ↩3 ↩4