A dataset created for the final project of the Speech Recognition Systems Evaluation Workshop (ZWESUI). The corpus focuses on the evaluation of ASR systems in the specific domain of houseplants and botany.

The dataset contains 523 audio segments originating from two main sources with different acoustic and linguistic characteristics:

Spontaneous speech (YouTube): Excerpts from educational and tutorial videos on plant cultivation.
Synthetic speech (TTS): Read speech generated from botanical texts using modern TTS engines (Coqui XTTS v2, KugelAudio-0-Open).

Data Breakdown

The corpus contains approximately 1 hour and 6 minutes (~66.6 minutes) of audio data:

Total Duration: ~3996 seconds
Spontaneous Speech (YouTube): 271 segments (~43.2 minutes)
Synthetic Speech (TTS Total): 252 segments (~23.4 minutes)
- KugelAudio-0-Open: 151 segments
- Coqui XTTS v2: 101 segments

Annotator Breakdown

The dataset was manually prepared and cross-verified by a 5-person team. Following the project requirements, each team member was responsible for transcribing/generating and cross-verifying an assigned pool of recordings:

Transcription / Generation (annotator):

s463015: 121 segments
511023: 102 segments
s512479: 100 segments
481875: 100 segments
s452376: 100 segments

Cross-validation (verified_by):

s512479: 121 segments
s452376: 102 segments
s463015: 100 segments
511023: 100 segments
481875: 100 segments

Tools & Methodology

The process of compiling and processing the dataset relied on the following tools and technologies:

Audio Generation (TTS): Coqui XTTS v2 and KugelAudio-0-Open engines.
Text Normalization: A custom Python pipeline (normalize.py) based on regular expressions, enforcing a uniform representation of numbers, abbreviations, and punctuation, fully documented in the normalization protocol.
Data Engineering: pandas and the HuggingFace datasets ecosystem were used to structure the audio files and transcriptions into the final format.
ASR Evaluation: Testing was conducted on infrastructure utilizing the MLX backend (Apple Silicon) for local execution of the Qwen3-ASR model, alongside standard transformers libraries for Whisper and API calls for Cohere.

Supported Tasks

automatic-speech-recognition: Evaluation, testing, and comparison of speech recognition systems for the Polish language, with an emphasis on specialized vocabulary (plant names, care treatments).

Dataset Structure

Data Fields

The dataset provides a standard column layout for the datasets library:

audio: Audio recording in WAV format (usually 16kHz, mono).
text: Original reference transcription.
text_norm: Transcription after a rigorous text normalization process (using rules prepared in the normalize.py file).
source: The origin source of the recording (e.g., youtube, coqui_xtts_v2, KugelAudio-0-Open).
duration: The length of the audio segment in seconds.
source_license: The license assigned to a given segment resulting from source rights (e.g., CC-BY 3.0 for YouTube, MIT for KugelAudio).

Evaluation & Baseline

The dataset was used to conduct a multidimensional evaluation of 3 ASR systems with different architectures. Results for the entire dataset (n=523) applying normalization for both references and hypotheses:

System	WER [%]	CER [%]	Bootstrap CI 95% (WER)
Whisper-large-v3	7.67	3.02	6.83 - 8.48
Cohere-Transcribe-03-2026	8.43	3.04	7.60 - 9.27
Qwen3-ASR-1.7B (MLX)	15.61	5.26	14.58 - 16.55

Quality Gap (Main observations): Error analysis revealed the largest quality gap between spontaneous and read speech. All systems (including Whisper-large-v3) show a significantly higher error rate (WER) for spontaneous speech segments from YouTube compared to highly regular, phonetically clean recordings from TTS engines.

Licensing

The corpus is a compilation of data from various sources, and each retains its original license:

Recordings obtained from YouTube are shared under the Creative Commons CC-BY 3.0 license (as the platform's standard Creative Commons license).
Generated synthetic recordings inherit the license of the respective models (MIT for KugelAudio-0-Open, Coqui Public Model License for Coqui XTTS v2).

Detailed license information for each of the 523 segments can be found in the source_license column.

Downloads last month: 21

Repository:

mszulcc/grupa-4-ZWESUI0

Language:

Polish (pl)

Domain:

Indoor plant cultivation (Houseplants / Botany)