README.md

---
license: other
license_name: ms-by-nc-nd
license_link: https://huggingface.co/datasets/TalTechNLP/EFAC/blob/main/LICENSE.pdf
language:
- et
extra_gated_prompt: >-
  You agree to the license terms of this dataset (no restribution, no commercial use)
extra_gated_fields:
  Company/university: text
  Website: text
  Planned usage of the dataset: text
task_categories:
- automatic-speech-recognition
- audio-classification
pretty_name: Estonian Foreign Accent Corpus
size_categories:
- 10K<n<100K
---
# Estonian Foreign Accent Corpus 1.0

The Estonian Foreign Accent Corpus (EFAC) is a corpus of Estonian speech
recordings by non-native speakers of Estonian, with a native Estonian reference
subset. It was created at the Laboratory of Language Technology, Tallinn
University of Technology, with funding from the national program *Estonian
Language Technology* (2006-2010; 2011-2017).

EFAC was developed for research on second-language pronunciation, foreign accent
characteristics, and speech technology applications such as automatic speech
recognition and pronunciation assessment.

## Speakers

The corpus contains speakers from a range of first-language backgrounds,
including Russian, Finnish, Latvian, German, Lithuanian, French, Japanese,
Swedish, Spanish, English, Italian, Hindi, Danish, Dutch, Slovak, Polish,
Portuguese, and Azerbaijani. A native Estonian reference subset of 10 male and
10 female speakers was also recorded.

Among the L2 speakers, approximately 60% are female and 40% are male. Speaker
ages range from 16 to 67 years. Most participants began learning Estonian
between ages 18 and 30, although some started as early as age 7 and others after
age 40. Participants had also studied multiple foreign languages, most commonly
English, German, or French.

## Recording Conditions

Most recordings were made in the TalTech recording studio. Some L2 speakers were
recorded at their home universities in Finland (Oulu, Helsinki, Turku), France
(Paris), Austria (Vienna), Latvia (Riga), and Lithuania (Vilnius).

Recordings were made with SpeechRecorder using two condenser microphones: a
close-talking microphone and a desktop microphone. Audio is stored as high
quality WAV files at 44.1 kHz, 16-bit stereo. Each speaker contributed
approximately 25 minutes of speech.

Approximately half of the read sentences are segmented and annotated at word and
phone levels in Praat.

## Contents

Each speaker contributed 143 speech items:

- Spontaneous speech:
  - self-introduction
  - 3 picture descriptions
- Read speech:
  - 137 isolated sentences covering major phonological features of Estonian,
    including vowel and consonant quantity oppositions, frequent diphthongs and
    consonant clusters, and palatalisation contrasts
  - 2 short stories
- Text prompts:
  - read-speech prompt texts are provided as UTF-8 encoded TXT files

## Directory Structure

The corpus is organized by first-language code and speaker:

```text
  AZ/
    VM/
      *.wav
  DE/
    BJ/
      *.wav
    BK/
      *.wav
  ...
```

Filenames in each speaker folder follow the format `xx_yy_zzz`, where `xx` is
the speaker ID, `yy` is the language code, and `zzz` is the prompt number. For
example, `db_en_002` represents speaker `db`, English as the speaker's L1, and
prompt number `002`.

This local copy also includes generated metadata files:

- `metadata.csv`: utterance-level metadata in Hugging Face AudioFolder style.
- `speaker_metadata.csv`: speaker-level metadata .

The `metadata.csv` columns are:

- `file_name`: relative path to the WAV file
- `ID`: utterance identifier
- `L1`: speaker first language as an ISO 639-3 code
- `speaker`: speaker identifier prefixed with the ISO 639-3 L1 code
- `gender`: speaker gender
- `proficiency`: Estonian proficiency level, with native Estonian speakers
  marked as `native`
- `text`: normalized transcript text

## Statistics

Corpus statistics:

| Measure | Value |
| --- | ---: |
| Utterances with metadata and transcripts | 31,835 |
| Speakers in metadata | 206 |
| First-language backgrounds in metadata | 18 |
| Indexed audio duration | 43.49 hours |
| Average indexed utterance duration | 4.92 seconds |
| Transcript tokens, whitespace-counted | 199,278 |

### Indexed utterances by first language

| L1 | Utterances | Speakers |
| --- | ---: | ---: |
| aze | 157 | 1 |
| dan | 311 | 2 |
| deu | 2,720 | 17 |
| eng | 899 | 5 |
| est | 3,116 | 21 |
| fin | 4,548 | 30 |
| fra | 2,006 | 13 |
| hin | 551 | 3 |
| ita | 882 | 6 |
| jpn | 1,020 | 7 |
| lav | 2,927 | 21 |
| lit | 1,709 | 12 |
| nld | 314 | 2 |
| pol | 343 | 2 |
| rus | 8,237 | 50 |
| slk | 291 | 2 |
| spa | 808 | 5 |
| swe | 996 | 7 |

## Licence and Access

The corpus available under the META-SHARE NonCommercial
NoRedistribution NoDerivatives (MS NC-NoReD-ND) license. 
You may use the corpus non-commercially, but you generally may not redistribute it or share copies onward, and you may not distribute derivatives.

## Citation

Please cite:

Meister, Einar; Meister, Lya (2015). Development and use of the Estonian L2
corpus. In Jürgen Trouvain (Ed.), *Book of Extended Abstracts. Workshop on
Phonetic Learner Corpora, 12 August 2015, Glasgow (Satellite Workshop of the
18th International Congress of Phonetic Sciences)*, 45-47.

```bibtex
@inproceedings{Meister2015EstL2Corpus,
  author    = {Meister, Einar and Meister, Lya},
  title     = {Development and Use of the Estonian L2 Corpus},
  booktitle = {Book of Extended Abstracts. Workshop on Phonetic Learner Corpora, 12 August 2015, Glasgow (Satellite Workshop of the 18th International Congress of Phonetic Sciences)},
  editor    = {Trouvain, Jürgen},
  pages     = {45--47},
  year      = {2015}
}
```