MEDHARVIX-SYSTEMS/bhasaflow-khasi-english-parallel-sample-v1

BhasaFlow Khasi-English Parallel Sample v1

A professionally curated, gold-standard parallel speech and text corpus for the Khasi language.

---

Published by Medharvix Systems Private Limited Part of the BhasaFlow Low-Resource Language Technology Initiative

---

Overview

This repository contains a public sample preview of the BhasaFlow Khasi-English Parallel Corpus, a structured speech and text dataset developed by Medharvix Systems Private Limited. The dataset pairs English source sentences with professionally translated Khasi text, each aligned with a corresponding studio-quality audio recording performed by native Khasi speakers.

This sample release is intended for evaluation, preliminary benchmarking, and collaboration discovery. It represents a carefully selected subset of a larger, production-grade corpus that is available through a controlled access process.

---

Why This Dataset Matters

Khasi (ISO 639-3: kha) is a language spoken by approximately 1.6 million people, primarily in the state of Meghalaya in Northeast India. Despite its cultural significance and active daily usage, Khasi remains severely underrepresented in modern speech and language technology. There are few publicly available parallel corpora, and almost no structured speech datasets suitable for training automatic speech recognition or text-to-speech systems.

The BhasaFlow Khasi-English Parallel Corpus directly addresses this gap. It provides researchers, developers, and institutional partners with a reliable, well-documented, and quality-controlled dataset that can serve as a foundation for:

• Building speech recognition systems for Khasi
• Developing text-to-speech synthesis in Khasi
• Training and evaluating machine translation models between English and Khasi
• Conducting linguistic research on Khasi phonology, morphology, and syntax
• Supporting digital language preservation and revitalization efforts

---

About Medharvix Systems Private Limited

Medharvix Systems Private Limited is an AI-focused technology company building structured language infrastructure for underserved linguistic communities. Our work spans data engineering, model development, and deployment of language technology systems designed for real-world adoption in education, governance, healthcare, and public communication.

About BhasaFlow

BhasaFlow is the language technology division of Medharvix Systems. It operates as a dedicated initiative focused on:

• Data collection and curation for low-resource languages of Northeast India
• Model training and benchmarking for speech and text processing
• Tool development for translation, transcription, and language access
• Partnership coordination with academic institutions, government bodies, and research organizations

BhasaFlow follows a structured pipeline approach: data is collected through organized contributor networks, reviewed through multi-stage quality assurance, and released in controlled phases to ensure integrity and responsible use.

---

Dataset Scope

Attribute	Detail
Language pair	Khasi (kha) and English (en)
Modality	Text (parallel) + Audio (speech)
Sample size (this release)	100 sentence pairs with aligned audio
Audio format	WAV, 16-bit PCM
Speaker gender distribution	Male and female speakers
Domain coverage	News, legal, governance, technology, religion, daily life, sports
Release type	Public preview / sample subset
Full corpus access	Permission required

---

Language Coverage

Language	ISO 639-3	Script	Region	Speaker Population
Khasi	kha	Latin	Meghalaya, India	~1.6 million
English	eng	Latin	Global	~1.5 billion

Khasi belongs to the Austroasiatic language family and is the official language of the Khasi Hills district in Meghalaya. It uses a Latin-based script adapted during the colonial period and remains the primary written form in education and media.

---

Data Collection Methodology

The dataset was produced through a structured, multi-phase workflow:

1. Source selection: English sentences were curated from diverse domains including news reporting, legal and governmental language, religious texts, everyday conversation, technology descriptions, and sports commentary. Sources were selected to ensure topical breadth and linguistic variety.

2. Translation: Each English sentence was translated into Khasi by native Khasi speakers with demonstrated fluency in both languages. Translators were briefed on accuracy standards, with emphasis on preserving meaning, register, and natural phrasing.

3. Recording: Audio recordings were produced by native Khasi speakers in controlled environments. Recordings follow consistent technical parameters (WAV format, 16-bit PCM) to ensure suitability for speech processing applications.

4. Alignment: Each audio file is explicitly linked to its corresponding text pair through a structured metadata file, enabling direct use in ASR, TTS, and translation pipelines.

---

Recording and Curation Process

All audio recordings in this dataset were produced under the following conditions:

• Speakers: Native Khasi speakers, both male and female, selected for clear articulation and natural speech patterns
• Environment: Controlled recording conditions to minimize background noise and ensure audio clarity
• Format: Uncompressed WAV files at standard speech processing specifications
• Review: Each recording was reviewed for completeness, clarity, and alignment with the corresponding text

Recordings that did not meet quality thresholds during review were re-recorded or excluded from the release.

---

Team Structure and Contributor Workflow

The BhasaFlow data collection effort was carried out through a coordinated team structure:

• 20 contributor teams participated in the broader data creation and review process
• Each team included native Khasi speakers responsible for translation, recording, or quality review
• Team leads coordinated task assignment, progress tracking, and output review
• A central quality assurance team performed cross-checks across all contributor outputs

This distributed workflow ensured both scale and consistency. Every data point in this release has been touched by multiple contributors across the pipeline.

---

Gold Data Quality Statement

This dataset is positioned as gold-standard curated data. This designation reflects the following characteristics:

• Every sentence pair has been translated by a qualified native speaker, not machine-generated
• Every audio recording has been produced by a native speaker in a controlled setting
• Every entry has passed through at least one manual review stage
• Text alignment between English and Khasi has been verified
• Audio-text alignment has been validated
• Entries that failed quality checks were excluded from the release

Medharvix Systems does not use the term "gold standard" loosely. It reflects a genuine investment in manual review, structured workflows, and contributor accountability.

---

Annotation and Validation Process

The annotation process followed a three-stage pipeline:

Stage 1: Initial Translation and Recording

Native Khasi contributors produced translations and audio recordings based on assigned English source sentences. Each contributor worked within defined quality guidelines covering translation accuracy, audio clarity, and metadata completeness.

Stage 2: Peer Review

A second contributor reviewed each translation for accuracy, naturalness, and completeness. Audio recordings were checked for clarity, correct pronunciation, and alignment with the written Khasi text.

Stage 3: Quality Assurance

A dedicated QA team performed final checks on the assembled dataset, including:

• Text consistency and encoding verification
• Audio integrity checks (no corruption, correct format)
• Metadata completeness and cross-referencing
• Duplicate detection and removal

---

Quality Assurance Pipeline

Source Selection
      |
      v
Translation by Native Speakers
      |
      v
Audio Recording (Controlled Environment)
      |
      v
Peer Review (Accuracy + Clarity)
      |
      v
Quality Assurance (Format + Alignment + Integrity)
      |
      v
Final Dataset Assembly
      |
      v
Release Staging and Review

---

Dataset Schema

Field	Type	Description
sentence_id	int	Unique numerical identifier for each sentence pair
english_text	string	The English source sentence
khasi_text	string	The corresponding Khasi translation, produced by a native speaker
gender	string	Gender of the audio speaker (male or female)
file_name	string	Filename of the corresponding WAV audio recording
audio	Audio	The speech audio file (WAV, 16-bit PCM)

All text fields use UTF-8 encoding. Audio files are uncompressed WAV format suitable for direct ingestion into speech processing pipelines.

---

Example Rows

The following rows are drawn from the actual dataset to illustrate format and content:

sentence_id	english_text	khasi_text	gender	file_name
272	Over 120 students took part in this session.	Palat 120 ngut ki samla pule ki la iashim bynta ha kane ka jingialang.	male	audio_272.wav
93	Isha Ambani and Anand Piramal are married.	Ka Isha Ambani bad u Anand Piramal ki la ia poi kha.	female	audio_93.wav
402	This is a huge reform.	Kane ka dei ka jingpynkylla kaba khraw.	female	audio_402.wav

---

Public Sample vs. Full Corpus

This repository contains a public sample preview of the BhasaFlow Khasi-English Parallel Corpus.

	Public Sample (this repo)	Full Corpus
Access	Open	Permission required
Size	100 sentence pairs	Extended corpus
Purpose	Evaluation, benchmarking, collaboration discovery	Research, development, production
Audio	Included	Included
Metadata	Full schema	Full schema
License	Restricted (see below)	Granted upon approval

The public sample is designed to allow researchers and developers to:

• Evaluate the dataset format and quality before requesting full access
• Run preliminary experiments and benchmarks
• Assess suitability for their specific use case
• Demonstrate alignment with their research or development objectives

The full corpus is not openly downloadable. Access is granted through a structured permission process.

---

Access and Permissions

Requesting Full Dataset Access

Access to the complete BhasaFlow Khasi-English Parallel Corpus is available to qualified researchers, academic institutions, government organizations, and enterprise partners through a permission-based process.

To request access:

1. Submit a request through the Hugging Face dataset access form (if gated access is enabled)
2. Contact us directly at contact@medharvix.com with a brief description of your intended use
3. Include your affiliation, project context, and expected application of the dataset

All requests are reviewed by the BhasaFlow team. Approved users receive access credentials and usage guidelines.

---

For access to the full BhasaFlow Khasi-English corpus, please request permission or contact contact@medharvix.com.

---

Alternative Contact Options

• Research and academic access: Please include your institutional affiliation and research summary
• Government and institutional partnerships: Contact us for tailored collaboration arrangements
• Enterprise and commercial licensing: We welcome discussions on integration and deployment use cases
• Strategic partnerships: For joint research, co-development, or regional language initiatives

---

Intended Use

This dataset is intended for the following applications:

• Automatic Speech Recognition (ASR): Training and evaluating speech-to-text models for Khasi
• Text-to-Speech (TTS): Developing speech synthesis systems for the Khasi language
• Machine Translation: Building English-to-Khasi and Khasi-to-English translation models
• Linguistic Research: Phonological, morphological, and syntactic analysis of Khasi
• Language Preservation: Supporting digital documentation and revitalization of Khasi
• Benchmarking: Establishing performance baselines for low-resource language technology

---

Out-of-Scope Use

This dataset should not be used for:

• Speaker identification, re-identification, or surveillance of any kind
• Voice cloning or synthetic impersonation without explicit contributor consent
• Generating misleading or deceptive content in Khasi or English
• Any application that could harm the Khasi-speaking community or its cultural heritage
• Redistribution or republication without authorization from Medharvix Systems

---

Ethical Considerations

• All contributors were native Khasi speakers who participated voluntarily in the data creation process
• Contributors were informed about the intended use of their recordings and translations
• The dataset does not contain personally identifiable information beyond speaker gender
• Audio recordings do not include names, addresses, or other identifying details of the speakers themselves
• Medharvix Systems is committed to ensuring that this data supports rather than exploits the Khasi-speaking community

---

Bias and Limitations

• Speaker diversity: While the dataset includes both male and female speakers, the current sample may not fully represent all dialectal variations within the Khasi language
• Domain coverage: The dataset covers multiple domains but is not exhaustive; certain specialized vocabularies may be underrepresented
• Translation variability: Khasi translations reflect the linguistic judgment of individual native speakers; alternative valid translations may exist for any given sentence
• Sample size: This public release contains 100 sentence pairs, which is sufficient for evaluation but may not support large-scale model training without augmentation
• Regional variation: Khasi has dialectal variation across different regions of Meghalaya; this dataset primarily reflects standard Khasi as used in education and media

Users should consider these limitations when designing experiments or drawing conclusions from results obtained using this dataset.

---

Licensing and Usage Conditions

This dataset is released under a restricted license managed by Medharvix Systems Private Limited.

Terms of use:

• The public sample may be used for non-commercial research, evaluation, and benchmarking
• Redistribution of the dataset (in whole or in part) is not permitted without written authorization
• Commercial use requires a separate licensing agreement with Medharvix Systems
• Attribution to Medharvix Systems Private Limited and the BhasaFlow initiative is required in all publications and outputs derived from this dataset
• Users must comply with all applicable data protection and privacy regulations

For licensing inquiries, contact contact@medharvix.com.

---

Responsible Use and Compliance

Medharvix Systems is committed to the responsible release and use of language technology resources. Users of this dataset are expected to:

• Use the data in accordance with the stated intended purposes
• Avoid applications that could cause harm to the Khasi-speaking community
• Report any issues, errors, or concerns to the BhasaFlow team
• Respect the intellectual property and cultural significance of the materials
• Obtain appropriate permissions before integrating this data into deployed systems

Fuller access to the complete corpus may be granted for approved use cases following review by the BhasaFlow data governance team.

---

Contact

Organization	Medharvix Systems Private Limited
Initiative	BhasaFlow
Email	contact@medharvix.com
Hugging Face	MEDHARVIX-SYSTEMS

For research collaborations, institutional partnerships, enterprise licensing, or general inquiries, please reach out to our team.

---

Citation

If you use this dataset in your research or applications, please cite it as follows:

@dataset{bhasaflow_khasi_english_v1_2026,
  title        = {BhasaFlow Khasi-English Parallel Sample v1},
  author       = {Medharvix Systems Private Limited},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/MEDHARVIX-SYSTEMS/bhasaflow-khasi-english-parallel-sample-v1},
  license      = {Restricted},
  language     = {kha, en},
  version      = {1.0}
}

---

Closing Note

The BhasaFlow Khasi-English Parallel Corpus represents a deliberate and sustained investment in language technology infrastructure for one of India's most important yet underserved linguistic communities. Every sentence pair, every audio recording, and every quality check in this dataset reflects the work of native Khasi contributors and a coordinated operational team.

Medharvix Systems Private Limited is open to collaboration with researchers, academic institutions, government stakeholders, private companies, and strategic partners who share our commitment to building reliable, high-quality language technology for the languages that need it most.

We invite qualified organizations and individuals to explore this sample, evaluate its quality, and reach out to discuss how we can work together.

---

Medharvix Systems Private Limited Building structured language AI infrastructure for underserved communities.

BhasaFlow — Precision language technology for the languages the world forgot.