suleiman2003/unified-hausa-speech

Unified Hausa Speech Dataset v5

Dataset Description

A large-scale, cleaned, deduplicated, and quality-filtered Hausa speech dataset compiled from 6 open-source collections. Designed for Text-to-Speech (TTS) and Automatic Speech Recognition (ASR) research on one of Africa's most widely spoken languages.

Hausa (ISO 639-1: ha) is a Chadic language spoken by over 80 million people across West and Central Africa — primarily in Nigeria and Niger, and as a trade language across the Sahel region. Despite its large speaker population, high-quality speech data remains scarce compared to global languages.

This dataset addresses that gap by unifying, cleaning, and standardising multiple existing Hausa speech collections into a single, ready-to-train resource.

Key Properties

• Audio: 16 kHz mono FLAC — silence-trimmed, loudness-normalised to −20 dBFS
• Text: NFC-normalised Hausa transcripts
• Speakers: Sequential integer IDs, stable across pipeline runs
• Sorted: All clips from the same speaker appear consecutively (useful for speaker-adaptive TTS and speaker embedding extraction)
• Deduplicated: Cross-source deduplication by audio hash and transcript text
• Quality-filtered: Clips filtered for duration, clipping, and empty audio

---

Dataset Summary

Metric	Train	Test	Total
Total Clips	782,297	8,827	791,124
Total Duration	697.84 hrs	8.96 hrs	706.8 hrs
Unique Speakers	2,059	51	2,110
Avg Clip Duration	3.211s	3.654s	—

---

Dataset Structure

Features

Column	Type	Description
audio	Audio (16 kHz)	Mono FLAC audio, silence-trimmed, loudness-normalised to −20 dBFS
text	string	NFC-normalised Hausa transcript (Unicode canonical decomposition)
speaker_id	int64	Unique sequential speaker identifier — stable across pipeline runs; speakers from different sources never collide
gender	string	"male", "female", or "other" (fallback when gender is unknown or not provided)
duration	float32	Clip duration in seconds (post-processing, after speech trimming)

Splits

Split	Sources	Description
train	WaxalNLP, NaijaVoices (3 batches), BibleTTS train, TWB Voice train, AfricanVoices (6 batches), Common Voice train	General-purpose training data — diverse speakers, recording conditions, and domains
test	BibleTTS test + dev, TWB Voice dev + test, Common Voice dev + test	Held-out evaluation data — never mixed into train; speakers may overlap with train in some sources

Sorting: The dataset is sorted by speaker_id within each split, so all clips from the same speaker appear consecutively. This is beneficial for:

• Speaker-adaptive TTS fine-tuning (e.g., VITS, StyleTTS2)
• Speaker embedding extraction (e.g., d-vector, x-vector)
• Speaker diarisation evaluation
• Batch construction with per-speaker sampling

---

Sources

This dataset aggregates the following Hausa speech collections. Each source was independently processed through the same cleaning pipeline.

Source	Repository	Clips Kept	Output Splits
WaxalNLP	google/WaxalNLP	1,435	train
NaijaVoices	naijavoices/naijavoices-dataset	308,100	train
BibleTTS	vpetukhov/bible_tts_hausa	22,057	train + test
TWB Voice	CLEAR-Global/TWB-Voice-1.0	7,032	train + test
AfricanVoices	https://africanvoices.io	30,237	train
Common Voice	Mozilla Common Voice 26.0	3,219	train + test

Source Descriptions

1. WaxalNLP (google/WaxalNLP): High-quality read speech produced by Google for West African language NLP research. Covers a range of topics and speakers.

2. NaijaVoices (naijavoices/naijavoices-dataset): Community-contributed Nigerian speech recordings. Includes diverse speakers across multiple Nigerian languages; this pipeline extracts only Hausa clips. Processed in 3 sequential batches.

3. BibleTTS (vpetukhov/bible_tts_hausa): Studio-quality single-speaker Bible readings in Hausa. Exceptionally clean audio with professional-grade recording conditions. The train split contributes to pipeline train; dev and test splits contribute to pipeline test.

4. TWB Voice (CLEAR-Global/TWB-Voice-1.0): Translators Without Borders (now CLEAR Global) voice datasets created for humanitarian communication. Contains domain-specific vocabulary related to health, crisis response, and community information.

5. AfricanVoices (africanvoices.io): Large-scale crowd-sourced African speech platform. The Hausa dataset was downloaded as 6 batches via the AfricanVoices API. Audio is FLAC with metadata including speaker ID, gender, age group, education level, domain, and SNR.

6. Common Voice (Mozilla Common Voice 26.0): Mozilla's crowd-sourced speech corpus. Contributors read prompted sentences via a web interface. Community-validated through an up/down voting system. Downloaded via the Mozilla Data Collective API.

---

Duration Analysis

Per-Split Statistics

Statistic	Train	Test
Min	1.000s	1.024s
Max	29.888s	29.792s
Mean	3.211s	3.654s
Median	2.624s	2.944s
Std	2.098s	2.515s

Duration Distribution (Train)

Range	Clips	Percentage
< 2 seconds	210,276	26.9%
2 — 5 seconds	473,249	60.5%
5 — 10 seconds	86,453	11.1%
10 — 20 seconds	11,367	1.5%
20 — 30 seconds	952	0.1%

Note: Clips shorter than 1.0s or longer than 30.0s were rejected during quality filtering (see Quality Filtering section below).

---

Speaker Analysis

Top 20 Speakers by Sample Count (Train)

Rank	Speaker ID	Gender	Clips	Total Duration	Avg Duration
1	1888	male	39,295	5071.9 min	7.74s
2	189	female	8,367	422.8 min	3.03s
3	139	female	8,361	421.5 min	3.03s
4	217	female	8,341	416.4 min	3.00s
5	69	female	8,335	405.1 min	2.92s
6	88	male	8,332	415.7 min	2.99s
7	225	male	8,328	513.9 min	3.70s
8	40	female	8,321	553.0 min	3.99s
9	191	female	8,320	457.5 min	3.30s
10	86	male	8,316	390.2 min	2.81s
11	247	male	8,299	377.4 min	2.73s
12	79	female	8,295	416.4 min	3.01s
13	141	male	8,235	373.1 min	2.72s
14	159	female	8,224	297.5 min	2.17s
15	11	female	7,619	319.9 min	2.52s
16	275	female	7,596	320.6 min	2.53s
17	55	female	7,526	360.4 min	2.87s
18	76	female	7,502	365.3 min	2.92s
19	134	female	7,339	326.9 min	2.67s
20	13	male	7,206	305.9 min	2.55s

Speaker IDs are sequential integers assigned during pipeline processing. Each source's speakers are namespaced to prevent collisions (e.g., a speaker from WaxalNLP will never share an ID with a speaker from Common Voice, even if the raw identifiers happen to match).

---

Gender Distribution

By Number of Speakers

Gender	Train Speakers	Test Speakers
Male	1111	18
Female	947	13
Other/Unknown	1	20

By Number of Clips

Gender	Train Clips	Test Clips
Male	412,849	3,115
Female	367,645	4,296
Other/Unknown	1,803	1,416

"Other/Unknown" includes clips where the source did not provide gender metadata, or where the reported gender did not match standard categories. This is common in crowd-sourced datasets where demographic fields are optional.

---

Text Analysis

Statistic	Train	Test
Min Length	2 chars	6 chars
Max Length	24311 chars	329 chars
Mean Length	45.5 chars	49.3 chars
Median Length	40 chars	42 chars

All transcripts are normalised using Unicode NFC form. Minimum transcript length for inclusion is 2 characters.

---

Quality Filtering Summary

During processing, every clip is evaluated against multiple quality criteria. Clips that fail any criterion are rejected and excluded from the final dataset.

Filter Reason	Rejected Clips	Description
No Speaker ID	0	Source did not provide a speaker identifier
Too Short (< 1.0s)	4,294	Clip too brief for reliable TTS/ASR training
Too Long (> 30.0s)	9,995	Exceeds maximum duration; may contain multiple utterances
Excessive Clipping (> 1%)	303	Audio amplitude at digital ceiling — distorted recording
Empty After Processing	0	No audio samples remained after silence trimming
Processing Errors	0	Decoding failure, corrupt file, or unexpected format
Total Rejected	14,592
Total Kept	372,080

Deduplication

Cross-source deduplication is applied using two strategies:

• Audio-level: SHA-256 hash of raw audio bytes — catches identical recordings uploaded to multiple sources
• Text-level: Normalised transcript hash — catches re-recordings of the same sentence (configurable, may keep both if speakers differ)

---

Audio Processing Pipeline

Every clip passes through the following processing steps (in order):

1. Resample to 16 kHz mono (if source sample rate differs)
2. Noise reduction via spectral gating (noisereduce library)
3. Silence trimming using energy-based detection (librosa.effects.trim at 25 dB threshold)
4. Loudness normalisation to −20 dBFS (peak-based scaling)
5. Quality checks: duration bounds, clipping ratio, empty audio detection
6. FLAC encoding at 16-bit PCM (lossless compression, ~50% size reduction vs WAV)

Audio Specifications

Parameter	Value
Sample Rate	16,000 Hz
Channels	Mono
Format	FLAC (16-bit PCM)
Loudness Target	−20 dBFS (peak)
Silence Trimming	25 dB energy threshold
Min Duration	1.0s
Max Duration	30.0s
Max Clipping Ratio	1%
Noise Reduction	Spectral gating (stationary noise)

---

Usage

Basic Loading

from datasets import load_dataset

# Load the full dataset
ds = load_dataset("suleiman2003/unified-hausa-speech")

# Load a specific split
train = load_dataset("suleiman2003/unified-hausa-speech", split="train")
test = load_dataset("suleiman2003/unified-hausa-speech", split="test")

# Access a sample
sample = train[0]
audio_array = sample["audio"]["array"]         # numpy float32 array
sample_rate = sample["audio"]["sampling_rate"]  # 16000
transcript  = sample["text"]                    # Hausa text
speaker     = sample["speaker_id"]              # int64
gender      = sample["gender"]                  # "male" / "female" / "other"
duration    = sample["duration"]                # float32, seconds

Streaming (Memory-Efficient)

ds_stream = load_dataset("suleiman2003/unified-hausa-speech", split="train", streaming=True)
for sample in ds_stream:
    # Process one sample at a time — dataset never fully loaded into RAM
    pass

Filter by Speaker

# Get all clips from speaker 1
speaker_1 = train.filter(lambda x: x["speaker_id"] == 1)
print(f"Speaker 1 has {len(speaker_1)} clips")

Filter by Gender

female_clips = train.filter(lambda x: x["gender"] == "female")
print(f"{len(female_clips)} female clips")

Use with Transformers (ASR Fine-Tuning)

from transformers import WhisperProcessor, WhisperForConditionalGeneration

processor = WhisperProcessor.from_pretrained("openai/whisper-small")
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small")

def preprocess(batch):
    audio = batch["audio"]
    inputs = processor(audio["array"], sampling_rate=audio["sampling_rate"],
                       return_tensors="pt")
    batch["input_features"] = inputs.input_features[0]
    batch["labels"] = processor.tokenizer(batch["text"]).input_ids
    return batch

train_processed = train.map(preprocess)

---

Known Limitations

1. Gender metadata quality: Many sources provide gender as optional or self-reported. The "other" category includes both non-binary identifications and unknown/missing values.

2. Speaker overlap across sources: The same physical speaker may appear in multiple source datasets under different IDs. Cross-source speaker linking is not performed.

3. Dialect variation: Hausa has multiple dialects (e.g., Kananci, Zazzaganci, Katsinanci, Bausanchi). Source datasets do not consistently label dialect, so this dataset mixes dialect variants.

4. Recording conditions: Audio quality varies significantly across sources — from studio recordings (BibleTTS) to crowd-sourced mobile recordings (Common Voice, AfricanVoices). Noise reduction is applied but cannot fully compensate for poor source quality.

5. Transcript accuracy: Transcripts are taken as-is from each source. No independent verification or correction has been applied. Some transcripts may contain errors, particularly in crowd-sourced collections.

6. No forced alignment: Audio-text alignment is not verified at the word level. Some clips may contain untranscribed speech or the transcript may not exactly match the spoken content.

---

Ethical Considerations

• All source datasets are publicly available under open licences.
• Speaker identifiers are anonymised sequential integers with no link to personal identity.
• This dataset should not be used to build systems that impersonate specific individuals without consent.
• Users should be mindful of potential biases in crowd-sourced speech data, including demographic imbalances and regional representation gaps.

---

Citation

If you use this dataset, please cite the original sources:

@misc{unified_hausa_speech_v5,
  title={Unified Hausa Speech Dataset v5},
  author={Suleiman Ismail},
  year={2025},
  url={https://huggingface.co/datasets/suleiman2003/unified-hausa-speech},
  note={Aggregated from WaxalNLP, NaijaVoices, BibleTTS, TWB Voice, AfricanVoices, and Common Voice}
}

---

License

Apache 2.0 — see individual source licences for additional terms.