Polish-TEDx-ASR-Eval

A dataset for evaluating automatic speech recognition (ASR) systems for Polish in the domain of TEDx public talks.

Contains audio segments from Polish TEDx talks available on YouTube (CC BY-NC-ND 4.0) and synthetic speech generated with KugelAudio (MIT), with manually created and cross-verified transcriptions. Created as part of the course "Workshops on Evaluation of Speech Recognition Systems" (ZWESUI, AMU 2026) by Group 1.

Statistics

Feature	Value
Total segments	344
- YouTube segments	294
- TTS segments (KugelAudio)	50
Total audio duration	2601 s (~43 min)
- YouTube	~39 min
- TTS	~4 min
Minimum segment length	3.1 s
Maximum segment length	18.3 s
Average segment length	7.6 s
Median segment length	6.9 s
Number of source recordings (YouTube)	16
Number of annotators	5
Sample rate	16 kHz mono
Format	WAV
Speech type	spontaneous, planned public presentations + synthetic
Acoustics	conference/theatre hall, reverberation, audience reactions

Sources

YouTube (CC BY-NC-ND 4.0)

All YouTube recordings come from Polish TEDx talks.

Audio file	Source
audio_grupa1_0t-sG8FhC4E.wav	https://www.youtube.com/watch?v=0t-sG8FhC4E
audio_s441306_2-_tJd9FFK0.wav	https://www.youtube.com/watch?v=2-_tJd9FFK0
audio_s441306_4XMKrqabpds.wav	https://www.youtube.com/watch?v=4XMKrqabpds
audio_s441306_DOIllGDNKw4.wav	https://www.youtube.com/watch?v=DOIllGDNKw4
audio_s485704_cuRVUT3bmek.wav	https://www.youtube.com/watch?v=cuRVUT3bmek
audio_s485704_owA_Z3navgg.wav	https://www.youtube.com/watch?v=owA_Z3navgg
audio_s485704_qaOljbJ-kSU.wav	https://www.youtube.com/watch?v=qaOljbJ-kSU
audio_s485936_MYH2qGScEVk.wav	https://youtu.be/MYH2qGScEVk
audio_s485936_eANk3-vRpCM.wav	https://youtu.be/eANk3-vRpCM
audio_s485936_f6W_8V7wFJA.wav	https://youtu.be/f6W_8V7wFJA
audio_s512752_7_OWTJtUK5k.wav	https://www.youtube.com/watch?v=7_OWTJtUK5k
audio_s512752_SIcY83qBTsQ.wav	https://www.youtube.com/watch?v=SIcY83qBTsQ
audio_s512752_pNzlSnMjIjE.wav	https://www.youtube.com/watch?v=pNzlSnMjIjE
audio_s512757_RF8eJ2iN2Rk.wav	https://www.youtube.com/watch?v=RF8eJ2iN2Rk
audio_s512757_sHY7TdnWFuQ.wav	https://www.youtube.com/watch?v=sHY7TdnWFuQ
audio_s512757_wDGAkp1sHM4.wav	https://www.youtube.com/watch?v=wDGAkp1sHM4

Synthetic TTS (KugelAudio - MIT)

50 segments generated with KugelAudio by annotator s485936. Files: s485936_tts_sample_000.wav – s485936_tts_sample_049.wav.

Speaker Distribution

Speaker list (YouTube recordings)

Speaker	Audio file	Gender	Age group	Acoustics	Topic	Segments
Radosław Kotarski	audio_grupa1_0t-sG8FhC4E.wav	M	30-39	reverberant_hall	education	27
Adam Wajrak	audio_s441306_2-_tJd9FFK0.wav	M	40-49	quiet_hall	ecology	12
Mariusz Czarnecki	audio_s441306_4XMKrqabpds.wav	M	40-49	reverberant_hall	business	32
Jan A.P. Kaczmarek	audio_s441306_DOIllGDNKw4.wav	M	50+	reverberant_hall	culture	10
Adam Haertle	audio_s485704_cuRVUT3bmek.wav	M	40-49	quiet_hall	cybersecurity	17
Piotr Zielonka	audio_s485704_owA_Z3navgg.wav	M	50+	reverberant_hall	psychology	17
Joanna Flis	audio_s485704_qaOljbJ-kSU.wav	F	30-39	reverberant_hall	psychology	17
Maciej Makselon	audio_s485936_MYH2qGScEVk.wav	M	30-39	reverberant_hall	linguistics	17
Magdalena Komsta	audio_s485936_eANk3-vRpCM.wav	F	30-39	reverberant_hall	medicine	17
Janina Bąk	audio_s485936_f6W_8V7wFJA.wav	F	30-39	reverberant_hall	psychology	17
Mateusz Kurleto	audio_s512752_7_OWTJtUK5k.wav	M	30-39	reverberant_hall	technology	20
Artur Małek	audio_s512752_SIcY83qBTsQ.wav	M	30-39	quiet_hall	psychology	20
Bogusław Pranszke	audio_s512752_pNzlSnMjIjE.wav	M	50+	quiet_hall	science	20
Róża Hajkuś	audio_s512757_RF8eJ2iN2Rk.wav	F	30-39	reverberant_hall	education	17
Kamil Kozieł	audio_s512757_sHY7TdnWFuQ.wav	M	30-39	reverberant_hall	education	17
Piotr Konieczny	audio_s512757_wDGAkp1sHM4.wav	M	30-39	reverberant_hall	cybersecurity	17

Gender distribution

Gender	Speakers	Segments	%
M	12	226	76.9%
F	4	68	23.1%

Age distribution

Age group	Speakers	Segments	%
30-39	10	186	63.3%
40-49	3	61	20.7%
50+	3	47	16.0%

Acoustic conditions

Conditions	Speakers	Segments	%
reverberant_hall	12	225	76.5%
quiet_hall	4	69	23.5%

Topic distribution

Topic	Segments	%
psychology	71	24.1%
education	61	20.7%
cybersecurity	34	11.6%
business	32	10.9%
technology	20	6.8%
science	20	6.8%
linguistics	17	5.8%
medicine	17	5.8%
ecology	12	4.1%
culture	10	3.4%

Data Structure

The file corpus.csv contains the following columns:

Column	Type	Description
file	string	WAV filename
start	float	Segment start time [s]
end	float	Segment end time [s]
text	string	Raw transcription (original spelling)
text_norm	string	Normalised transcription (ready for WER computation)
duration	float	Segment duration [s]
annotator	string	Username of the transcription author
verified_by	string	Username of the verifier
source_url	string	Source recording URL
source_license	string	Source recording licence

Example Entries

{
  "file": "audio_grupa1_0t-sG8FhC4E.wav",
  "start": 0.436,
  "end": 5.405,
  "text": "w ogóle być w stanie proponować jakiekolwiek alternatywne systemowe rozwiązania",
  "text_norm": "w ogóle być w stanie proponować jakiekolwiek alternatywne systemowe rozwiązania",
  "duration": 4.97,
  "annotator": "s485704",
  "verified_by": "s512752",
  "source_url": "https://www.youtube.com/watch?v=0t-sG8FhC4E",
  "source_license": "CC BY-NC-ND 4.0 (TEDx Talks)"
}

{
  "file": "audio_s512757_wDGAkp1sHM4.wav",
  "start": 834.307,
  "end": 839.642,
  "text": "wiele osób się obawia i zakleja kamerę czy jest na sali ktoś kto zakleja kamerę na na swoim laptopie ręka w górę",
  "text_norm": "wiele osób się obawia i zakleja kamerę czy jest na sali ktoś kto zakleja kamerę na na swoim laptopie ręka w górę",
  "duration": 5.335,
  "annotator": "s512757",
  "verified_by": "",
  "source_url": "https://www.youtube.com/watch?v=wDGAkp1sHM4",
  "source_license": "CC BY-NC-ND 4.0 (TEDx Talks)"
}

{
  "file": "s485936_tts_sample_000.wav",
  "start": 0.0,
  "end": 3.867,
  "text": "Zostawiam was dzisiaj z tym jednym pytaniem: kim będziecie, kiedy zgaśnie to sceniczne światło?",
  "text_norm": "zostawiam was dzisiaj z tym jednym pytaniem kim będziecie kiedy zgaśnie to sceniczne światło",
  "duration": 3.867,
  "annotator": "s485936",
  "verified_by": "",
  "source_url": "",
  "source_license": "MIT (KugelAudio)"
}

Annotators

Annotator	YouTube segments	TTS segments	Total
s441306	59	0	59
s485704	56	0	56
s485936	57	50	107
s512752	66	0	66
s512757	56	0	56

Transcription Methodology

• Who transcribed: 5 student annotators working independently on assigned source recordings
• Guidelines: Team normalisation protocol agreed before transcription began; rules cover lowercase conversion, punctuation removal, hesitation removal, loanword normalisation, and ordinal numeral expansion
• Verification: Cross-verification - each segment reviewed by a second annotator (field verified_by)
• Inter-annotator agreement: Not formally measured.
• Tools: Audacity (label tracks exported as .txt), ffmpeg (audio conversion to 16 kHz mono WAV), KugelAudio (TTS synthesis)

Normalisation Protocol

Transcriptions in text_norm were processed by normalize() from normalize.py. Rules applied in order:

1. R7 - ordinal numerals expanded (1. → pierwszy)
2. R1 - lowercase conversion
3. R6 - loanword normalisation (email → e-mail, wifi → wi-fi)
4. R3 - hesitation removal (ee, yy, hm, mm)
5. R2 - punctuation and symbol removal
6. Whitespace collapse

Licence

YouTube source recordings: CC BY-NC-ND 4.0 (TEDx Talks / YouTube).
KugelAudio TTS audio: MIT (KugelAudio).
Transcriptions and metadata: CC BY-NC-ND 4.0.