Datasets:
File size: 7,822 Bytes
beeff6e 8e1f0e9 beeff6e 8e1f0e9 beeff6e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 | ---
language:
- ny
tags:
- chichewa
- audio
- text-to-speech
- automatic-speech-recognition
pretty_name: Chichewa Speech Corpus for Text-to-Speech and ASR
---
# Dataset Card for the Chichewa Speech Corpus
The Chichewa Speech Corpus is a collection of read speech and corresponding text created for doctoral research on Chichewa text-to-speech (TTS) synthesis at the University of Botswana. The full research corpus contains **5,658 utterances totalling 40 hours and 12 minutes**, recorded by five native Chichewa speakers working as professional broadcasters in Malawi. The audio–text pairs can also support exploratory automatic speech recognition (ASR) research on read Chichewa.
## Dataset Details
### Dataset Description
The corpus contains human-recorded Chichewa speech paired with the corresponding written text. It captures a standardised reading style used by professional Malawian broadcasters. The speakers came from different parts of Malawi, and the prompts were selected from several sources and subject domains. The corpus was developed to train and evaluate a rule-guided neural TTS system for Chichewa.
- **Curated by:** Stanley Ndebvu, as part of doctoral research at the University of Botswana
- **Language(s) (NLP):** Chichewa (`ny`), as spoken by the participating Malawian broadcaster
- **Access and use terms:** [Read the Data Access and Use Terms](https://huggingface.co/datasets/Chithekitale/Chichewa_TTS_ASR/blob/main/DATA_USE_TERMS.md)
### Dataset Sources
- **Repository:** 10.57967/hf/11726
- **Research reference:** Stanley Ndebvu, *Advancing Naturalness in Low-Resource Speech Synthesis through Rule-Guided Neural Architectures: A Chichewa Case Study*, University of Botswana. [Insert thesis repository URL when available.]
## Uses
### Direct Use
The corpus was designed for training and evaluating Chichewa TTS systems. The paired audio and text can also be used for exploratory ASR experiments on read speech and for analysing pronunciation, speaking rate, and prosody in the recorded speaking style.
### Out-of-Scope Use
The corpus is unsuitable as the sole basis for claims about performance across all Chichewa speakers, dialects, or speaking styles. Evaluation of conversational or noisy speech requires additional data from those settings. Speaker impersonation, deceptive voice cloning, and attempts to identify the speakers are outside the intended uses of this corpus.
## Dataset Structure
The basic unit is an utterance paired with its corresponding Chichewa text. The thesis documents mono recordings captured at **16 kHz in 16-bit PCM**. The full corpus contains **5,658 utterances** and **40 hours and 12 minutes** of speech from **a single speaker**.
| Item | Description |
|---|---|
| Audio | A recording of one read utterance. [WAV.] |
| raw_text | The Chichewa text corresponding to the recording. |
| normalised_text | The standardised Chichewa text corresponding to the audio recording. Where applicable, numbers and abbreviations are expanded into their spoken forms |
| Gender | Gender of the speaker |
## Dataset Creation
### Curation Rationale
The corpus was created to provide paired Chichewa speech and text for developing and evaluating a rule-guided neural TTS system. Prompts from multiple subject areas were selected to capture a range of vocabulary and pronunciation patterns, while a consistent reading style and recording setup supported acoustic modelling.
### Source Data
#### Data Collection and Processing
The research team selected written Chichewa prompts from multiple sources and subject domains. The prompts were screened for ambiguous spellings, excessive numbers of named entities, and strongly domain-specific expressions. The speaker read the selected prompts in indoor recording sessions supported by a professional audio producer. A condenser microphone was positioned at a fixed distance from each speaker, and recording levels were monitored to avoid clipping. The speaker was asked to read at a steady pace, articulate clearly, and use natural intonation. Sentences were re-recorded when errors occurred. Some sessions had occasional background noise, which was addressed through noise reduction and filtering during preprocessing.
#### Who are the source data producers?
A native Chichewa speakers working as professional broadcasters. The speaker was selected purposively for native fluency and clear articulation. The recorded speech reflects their professional read-speech style. The thesis documents speakers’ regions of origin, which provide background about the sample rather than verified dialect labels.
### Annotations
#### Annotation process
The speech–text pairing was based on the written prompts that the speakers read. Recording errors were addressed by re-recording the affected sentences. Any additional transcript-level checks, timing annotations, or linguistic labels should be described here if they are included in the release.
#### Who are the annotators?
The speaker read the prompts, and a professional audio producer supported the recording process.
#### Personal and Sensitive Information
The audio contains natural human voices that may remain recognisable when names and other direct identifiers are removed. The prompts were screened to limit excessive use of named entities.
## Bias, Risks, and Limitations
The corpus represents a purposively selected professional broadcaster reading prepared text in indoor recording sessions. Their speech reflects a standardised Malawian broadcast reading style, and the sample provides limited coverage of the pronunciation, tone, prosody, and vocabulary used by other Chichewa speakers. Performance on this corpus may differ from performance on spontaneous conversation, informal speech, speech recorded in noisier settings, or regional varieties within and beyond Malawi. Any ASR results obtained using this corpus should be reported in the context of read speech and the evaluation split used.
### Recommendations
Use a speaker-disjoint test set when evaluating generalisation to unseen speakers, and report whether the training and test sets share speakers. Evaluate on additional recordings from the target speaker groups, language varieties, and recording conditions before making broader claims. Follow the access and reuse terms in the repository record.
## Citation
Cite the exact dataset version and DOI used in your work. Cite the doctoral thesis when discussing the data collection methods or the original TTS experiments. Replace the placeholders below with the published repository metadata.
**BibTeX**
```bibtex
@misc{ndebvu_chichewa_speech_corpus,
author = {Ndebvu, Stanley},
title = {Chichewa Speech Corpus for Text-to-Speech and ASR},
year = {2026},
version = {1.0},
publisher = {Chithekitale/Chichewa_TTS_ASR},
doi = {10.57967/hf/11726},
url = {https://huggingface.co/datasets/Chithekitale/Chichewa_TTS_ASR},
note = {Data set}
}
```
**APA**
Ndebvu, S. (2026). *Chichewa speech corpus for text-to-speech and ASR* (Version [1,0) [Data set]. [Chithekitale/Chichewa_TTS_ASR]. https://doi.org/10.57967/hf/11726
## Glossary
- **TTS (text-to-speech):** Generation of spoken audio from written text.
- **ASR (automatic speech recognition):** Conversion of spoken audio into written text.
- **Utterance:** An individual recorded speech segment paired with its corresponding written text.
- **PCM (pulse-code modulation):** The digital audio encoding used to capture the recordings.
## More Information
Sections 4.3.2 and 4.3.5–4.3.9 of the doctoral thesis describe the speech dataset, recording constraints, and plans for data sharing.
## Dataset Card Authors
Stanley Ndebvu
## Dataset Card Contact
[ndebvu.s@mzuni.ac.mw +265 995 164 604] |