mohdali1/baltivoice-asr

BaltiVoice ASR Dataset

Dataset Description

BaltiVoice is the first publicly available Automatic Speech Recognition (ASR) dataset for the Balti language (ISO 639-3: bft), a critically low-resource Tibetic language spoken primarily in the Gilgit-Baltistan region of Pakistan and parts of Ladakh, India.

This dataset was collected, validated, and processed to build the first open-source Balti ASR system using OpenAI Whisper fine-tuning. The train and validation splits are strictly speaker-disjoint, ensuring no speaker appears in both sets, providing a rigorous benchmark for low-resource ASR.

---

Language

Property	Detail
Language	Balti (بلتی)
ISO Code	bft
Language Family	Sino-Tibetan → Tibeto-Burman → Tibetic
Script	Nastaliq (Arabic-based, RTL)
Region	Gilgit-Baltistan, Pakistan; Ladakh, India
Speakers	~400,000 (estimated)
Resource Level	Critically low-resource

Balti has very limited digital presence, almost no NLP tooling, and until this release, no publicly available ASR dataset.

---

Dataset Statistics

Split	Samples	Speakers	Estimated Hours
Train	9,519	122	~15.9 hours
Validation	538	14	~0.9 hours
Total	10,060	136	~16.8 hours

Splits are speaker-disjoint — no speaker overlap between train and validation sets (verified, overlap = 0).

Speaker Demographics

Property	Value
Total unique speakers	136
Gender (female)	1,246 utterances
Gender (undisclosed)	1,030 utterances
Age (twenties)	3,758 utterances
Age (thirties)	3,697 utterances
Age (teens)	97 utterances

Audio Properties

Property	Value
Format	WAV (16kHz, mono)
Avg Duration	~6.0 seconds
Min Duration	~1.0 seconds
Max Duration	~15.0 seconds
Sample Rate	16,000 Hz

Text Properties

Property	Value
Avg Words/Sentence	10.12
Avg Characters	48.80
Script	Nastaliq (RTL)

---

Dataset Structure

Each sample contains:

• audio: 16kHz mono WAV audio array
• sentence: Balti transcription in Nastaliq script

{
  "audio": {
    "array": [...],          # numpy array
    "sampling_rate": 16000
  },
  "sentence": "بوا لہ سلام بے اِنپا سلام سہ مہ بیاس"
}

---

Source & Collection

• Base data sourced from Mozilla Common Voice Balti (bft) contribution project.
• Only validated recordings were used (10,060 of 10,547 total clips).
• Splits created using speaker-disjoint sampling (GroupShuffleSplit on client_id metadata, seed 42).
• Manual inspection of 50 random validation samples yielded a 96% annotation agreement rate.

---

Usage

Load the dataset

from datasets import load_dataset, Audio

dataset = load_dataset("mohdali1/baltivoice-asr")
dataset = dataset.cast_column("audio", Audio(sampling_rate=16000))

print(dataset)
# DatasetDict({
#     train: Dataset({features: ['audio', 'sentence'], num_rows: 9519})
#     validation: Dataset({features: ['audio', 'sentence'], num_rows: 538})
# })

Preview a sample

sample = dataset["validation"][0]
print("Transcription:", sample["sentence"])
print("Audio shape:", sample["audio"]["array"].shape)
print("Sample rate:", sample["audio"]["sampling_rate"])

---

Model Trained on This Dataset

A fine-tuned Whisper model trained on BaltiVoice is available at:

👉 mohdali1/whisper-small-balti

Model	WER (%)	CER (%)
Whisper-small (zero-shot)	159.19	152.52
Whisper-base (fine-tuned)	44.54	15.61
Whisper-small (fine-tuned)	26.74	8.67

Zero-shot WER above 100% indicates hallucination — the model generates words not present in the reference. Fine-tuning reduces this to an impressive 26.74% WER and 8.67% CER on the 538-utterance speaker-disjoint validation set.

from transformers import pipeline

asr = pipeline(
    "automatic-speech-recognition",
    model="mohdali1/whisper-small-balti",
    generate_kwargs={"language": "urdu", "task": "transcribe"}
)
result = asr("your_balti_audio.wav")
print(result["text"])

---

Paper

This dataset and the associated ASR system are described in detail in:

BaltiVoice: A Speech Corpus and Fine-tuned Whisper ASR System for the Balti Language
Muhammad Ali, Independent Researcher, Gilgit-Baltistan, Pakistan
📄 arXiv: arXiv:2606.03504

---

Social Impact

Balti is an endangered language with very limited computational resources. This dataset contributes toward:

• Preserving the Balti language digitally.
• Enabling voice technology and dictation tools for Balti speakers.
• Supporting NLP research for low-resource Tibetic languages.
• Providing a rigorous, reproducible foundation for future Balti TTS, NER, and MT systems.

---

Limitations

• Audio collected from volunteer contributors — some variation in recording quality and background noise exists.
• Vocabulary is limited to read-speech domains; conversational WER would likely be higher.
• Model fine-tuned using the Urdu tokenizer as a proxy (closest supported Nastaliq script in Whisper).
• No Unicode normalization was applied; Arabic/Farsi Yeh ambiguity may affect exact WER calculations.

---

License

Creative Commons Attribution 4.0 (CC-BY-4.0)

---

Citation

If you use this dataset or the associated models in your research, please cite:

@misc{ali2026baltivoice,
  author    = {Muhammad Ali},
  title     = {BaltiVoice: A Speech Corpus and Fine-tuned Whisper ASR System for the Balti Language},
  year      = {2026},
  eprint    = {2606.03504},
  archivePrefix = {arXiv},
  primaryClass = {cs.CL},
  url       = {https://arxiv.org/abs/2606.03504}
}

---

Author

Muhammad Ali
Independent Researcher, Gilgit-Baltistan, Pakistan
Alumnus, The Islamia University of Bahawalpur (IUB)

🔗 HuggingFace · GitHub · ORCID · LinkedIn