ssws3/High-Quality-Indic-Speech-Datasets

Indic Language Speech Dataset Samples (High Quality)

A. SPEECH DATASETS -

• Languages: Hindi, Bengali, Marathi, Telugu, Tamil
• Audio Format: WAV, 16 kHz, Mono, 16-bit PCM
• Total Samples

Category	Samples	Description
1. Monolingual Indic	500	Single Indian language only (no English mixing). Informal, conversational, and short utterances included.
2. Code-Mixed	500	Natural intra-utterance language switching between English and Indian language. Conversational and dictation-style speech.
3. Accented English	100	English speech produced by speakers with diverse Indian regional accents.
4. Utterance Extremes	100	Dedicated samples for very short and very long speech durations.
5. Numeric-Dense	100	Speech containing sequences of numbers such as phone numbers, OTPs, account numbers, or IDs and Repeated or consecutive digits (e.g., "zero zero zero").
6. Terminology	100	Speech containing domain-specific keywords and technical terms that are difficult to transcribe without context.
Total	1400

B. OTHER DATASETS -

Category	Description
1. Native Indic "Chain of Thought" (CoT) Reasoning	Step-by-step solutions to complex problems (Math, Science, Logic) written originally in Indian languages by human experts—not translated from English.
2. Agentic Browser Use (Indic Language Navigation)	Recordings of humans completing real-world tasks on Indian websites (government portals, e-commerce, banking) that use regional languages.
3. Code-Switching Technical Reasoning (Hinglish/Tanglish)	Technical troubleshooting or logical debates where the language naturally switches between English and an Indian language (e.g., "Hinglish").
4. Indian Legal & Bureaucratic Reasoning	Analysis of Indian High Court judgments, land records, or tax laws in regional languages, annotated by legal experts.
5. Indigenous Medical Knowledge (Ayurveda/Siddha)	Structured reasoning data from traditional Indian medical systems (Ayurveda, Siddha, Unani) in their native languages (Sanskrit, Tamil, Urdu).
6. Hyper-Local Multimodal Understanding (Video + Causality)	Videos of Indian-specific physical tasks (e.g., weaving, pottery, street cooking, fixing a puncture) with expert narration explaining why actions are taken.
7. Dialect & Low-Resource Language Corpora	High-quality data for languages/dialects that have almost zero internet presence (Gondi, Santhali, Maithili, Tulu).
8. Native Indic Creative Writing & Nuance	Poems, satire, scripts, and lyrics written originally in regional languages, annotated for their "rasa" (emotional essence) and cultural metaphors.

Category 1: Monolingual Indic Speech

• Description: Single Indian language only (no English mixing). Informal, conversational, and short utterances included.
• Samples: 500 (100 per language, 50% male + 50% female)
• Format: Single-speaker
• Directory:

/samples/monolingual-lndic-speech/
├── 00001.wav
├── 00001.json
├── 00002.wav
├── 00002.json
...

• JSON Schema:

{
  "_id": "00001",
  "duration": 7.11,
  "audio": "/monolingual-lndic-speech/00001.wav",
  "transcription": "यह छत किसी बैकरी की है जिसमें की ऊपर की तरफ सजावट करने के लिए हार्ट वाले गुब्बारों का प्रयोग किया गया",
  "metadata": {
    "gender": "male",
    "language": "Hindi",
    "script": "Devanagari",
    "accent": "Uttar Pradesh Hindi",
    "bit_depth": "16-bit PCM",
    "sample_rate": 16000,
    "tag": "monolingual-lndic-speech"
  }
}

Category 2: Code-Mixed Speech

• Description: Natural intra-utterance language switching between English and Indian language. Conversational and dictation-style speech.
• Example: "Main kal office जाऊंगा but meeting cancel हो गई".
• Script: Devanagari OR Latin (both acceptable).
• Samples: 500 (100 per language, 50% male + 50% female)
• Format: Single-speaker
• Directory:

/samples/code-mixed-speech/
├── 00001.wav
├── 00001.json
├── 00002.wav
├── 00002.json
...

• JSON Schema:

{
  "_id": "00005",
  "duration": 4.41,
  "transcription": "car वगैरा मिलती है, service new car अगर किसी को लेनी हो तो यहाँ से मिल जाएगी.",
  "audio": "/code-mixed-speech/00005.wav",
  "metadata": {
    "gender": "male",
    "language": "Hindi",
    "accent": "Hindi UP Accent",
    "bit_depth": "16-bit PCM",
    "sample_rate": 16000,
    "tag": "code-mixed-speech"
  },
}

Category 3: English with Indian Regional Accents

• Description: English speech produced by speakers with diverse Indian regional accents.
• Accent Examples: South Indian English, Bengali-accented English, Punjabi-accented English
• Samples: 100 total (20 samples of each languages)
• Directory (Single-Speaker):

/samples/english-with-indian-regional-accents/
├── 00001.wav
├── 00001.json
├── 00002.wav
├── 00002.json
...

• JSON Schema (Single-Speaker):

{
  "_id": "00001",
  "audio": "/english-with-indian-regional-accents/00001.wav",
  "duration": 11.76,
  "transcription": "A combination of Canadian capital, quickly organized, and petitioned, for the same privileges.",
  "metadata": {
    "language": "english",
    "accent": "Gujarati English accent",
    "bit_depth": "16-bit PCM",
    "sample_rate": 16000,
    "tag": "english-with-indian-regional-accents",
    "gender": "male"
  }
}

Category 4: Utterance Length Extremes

• Description: Dedicated samples for very short and very long speech durations
• Short: Commands, fragments (e.g., "Okay", "Haan", "Next slide", "रुको") — under 5 seconds
• Long: Audio >2 minutes, storytelling or monologue
• Samples: 100 total
  • 50 very short (<5 seconds, 10 samples of each language)
  • 50 very long (>2 minutes, 10 samples of each language)
• Format: Single-speaker & Multi-speaker
• Language: Monolingual Indic OR Code-Mixed
• Directory:

/samples/utterance-length-extremes/
├── short/
│   ├── 00001.wav
│   ├── 00001.json
│   ├── 00002.wav
│   ├── 00002.json
│   ├── ...
├── long/
│   ├── 00001.wav
│   ├── 00001.json
│   ├── 00002.wav
│   ├── 00002.json
│   ├── ...

• JSON Schema (Short):

{
  "_id": "00001",
  "duration": 1.696,
  "audio": "/short-utterance-length-extremes/00001.wav",
  "transcription": "कूदो",
  "metadata": {
    "language": "hindi",
    "gender": "Male",
    "topic": "Keywords Spotting",
    "bit_depth": "16-bit PCM",
    "sample_rate": 16000,
    "tag": "short-utterance-length-extremes",
    "type": "conversational"
  }
}

• JSON Schema (Long):

{
  "_id": "00024",
  "audio": "/long-utterance-length-extremes/00024.wav"
  "duration": 88,
  "num_speakers": 2,
  "turns": [
    {
      "speaker_id": "1",
      "start_time": 0,
      "end_time": 3.12,
      "transcription": "अरे तेरा phone कहां से लिया? बड़ा cute लग रहा है."
    },
    {
      "speaker_id": "2",
      "start_time": 3.12,
      "end_time": 7.04,
      "transcription": "Oh, यह वाला? मैंने cinema post से लिया था, ले देखना है क्या?"
    },
    .
    .
    .
    {
      "speaker_id": "1",
      "start_time": 87.65,
      "end_time": 88.37,
      "transcription": "पसंद"
    }
  ],
  "metadata": {
    "language": "hindi",
    "topic": "रिश्ते",
    "bit_depth": "16-bit PCM",
    "sample_rate": 16000,
    "tag": "long-utterance-length-extremes",
    "type": "conversational"
  },
}

Category 5: Numeric-Dense Speech

• Description: Speech containing sequences of numbers such as phone numbers, OTPs, account numbers, or IDs and Repeated or consecutive digits (e.g., "zero zero zero").
• Example: "Mera number hai nine nine nine six five four three two one zero".
• Samples: 100 (20 samples of each language)
• Format: Single-speaker
• Language: Monolingual Indic OR Code-Mixed
• Directory:

/samples/numeric-dense-speech/
├── 00001.wav
├── 00001.json
├── 00002.wav
├── 00002.json
...

• JSON Schema:

{
  "_id": "00027",
  "duration": 10.48,
  "audio": "/numeric-dense-speech/00027.wav",
  "transcription": "नमस्ते मेरा नया एड्रेस सेक्टर 4 5 में स्थित है और मेरा फ्लैट नंबर 1 0 2 9 8 7 3 है नोट करें",
  "metadata": {
    "language": "hindi",
    "topic": "Address and flat details",
    "domain": "Address",
    "numeric_content": [
      "4 5",
      "1 0 2 9 8 7 3"
    ],
    "bit_depth": "16-bit PCM",
    "sample_rate": 16000,
    "tag": "numeric-dense-speech"
  },
  "verbatim": [
    {
      "word": "नमस्ते",
      "start_time": 0,
      "end_time": 0.4
    },
    {
      "word": "मेरा",
      "start_time": 0.88,
      "end_time": 1.52
    },
    {
      "word": "नया",
      "start_time": 1.57,
      "end_time": 1.76
    },
    .
    .
    .
    {
      "word": "करें",
      "start_time": 9.97,
      "end_time": 10.47
    }
  ]
}

Category 6: Keywords & Special Terminology

• Description: Speech containing domain-specific keywords and technical terms that are difficult to transcribe without context.
• Domain Examples:
  • Medical: "paracetamol 500mg", "blood pressure"
  • Legal: "Section 144", "FIR number"
  • Technical: Brand names, acronyms
• Samples: 100
• Format: Single-speaker
• Language: Monolingual Indic OR Code-Mixed
• Directory:

/samples/keywords-and-special-terminology/
├── 00001.wav
├── 00001.json
├── 00002.wav
├── 00002.json
...

• JSON Schema:

{
  "_id": "00007",
  "duration": 3.28,
  "audio": "/keywords-and-special-terminology/00007.wav",
  "transcription": "এই phone-এর RAM এবং Processor খুব ভালো performance দিচ্ছে।",
  "metadata": {
    "language": "bengali",
    "topic": "Smartphone specification performance",
    "domain": "Technical",
    "terminology": [
      "RAM",
      "Processor",
      "performance"
    ],
    "bit_depth": "16-bit PCM",
    "sample_rate": 16000,
    "tag": "keywords-and-special-terminology"
  },
  "verbatim": [
    {
      "word": "এই",
      "start_time": 0,
      "end_time": 0.28
    },
    {
      "word": "phone-এর",
      "start_time": 0.3,
      "end_time": 0.56
    },
    .
    .
    .
    {
      "word": "দিচ্ছে।",
      "start_time": 2.9,
      "end_time": 3.16
    }
  ]
}

OTHER CATEGORIES (It's Sample is not present in this HF Dataset Repo):

1. Native Indic "Chain of Thought" (CoT) Reasoning

• Description: Step-by-step solutions to complex problems (Math, Science, Logic) written originally in Indian languages by human experts—not translated from English.
• Why it's Critical: Current models can solve calculus in English but fail in Hindi because they don't know the "reasoning vocabulary" of Indian languages. They need to learn how a Tamil or Bengali speaker thinks through a problem logically.
• Capabilities: Enables models to perform "reasoning" (o1-level capability) in regional languages, crucial for education and scientific research in India.
• Example Usage: A Physics teacher solving a JEE Advanced problem in Hindi, writing out every intermediate step and law used.
• Schema Example:

  {
    "id": "hindi_phy_001",
    "language": "Hindi",
    "problem": "एक 5kg का ब्लॉक...",
    "reasoning_trace": [
      {"step": 1, "thought": "पहले हम न्यूटन के दूसरे नियम (F=ma) का उपयोग करेंगे...", "calculation": "5 * 9.8"},
      {"step": 2, "thought": "घर्षण बल (Friction) की गणना...", "calculation": "0.5 * 49"}
    ],
    "final_answer": "24.5 Newtons",
    "verified_by": "PhD_Physics_Expert"
  }
  

2. Agentic Browser Use (Indic Language Navigation)

• Description: Recordings of humans completing real-world tasks on Indian websites (government portals, e-commerce, banking) that use regional languages.
• Why it's Critical: Building "Agents" that can use computers. US data doesn't teach an agent how to navigate the complex, unique UIs of Indian sites like IRCTC, GST portal, or vernacular news apps.
• Capabilities: Enables AI agents to book tickets, file taxes, or shop online for users who only speak regional languages.
• Example Usage: A screen recording + action log of a user booking a train ticket on IRCTC using the Hindi interface.
• Schema Example:

  {
    "task_id": "irctc_book_hindi",
    "goal": "Book ticket from Delhi to Patna",
    "language": "Hindi",
    "actions": [
      {"step": 1, "action": "click", "element": "button_login_hindi", "url": "irctc.co.in"},
      {"step": 2, "action": "type", "text": "New Delhi", "field": "origin_station"},
      {"step": 3, "action": "scroll", "direction": "down", "reason": "Finding 'Tatkal' option"}
    ]
  }
  

3. Code-Switching Technical Reasoning (Hinglish/Tanglish)

• Description: Technical troubleshooting or logical debates where the language naturally switches between English and an Indian language (e.g., "Hinglish").
• Why it's Critical: 90% of Indian internet users don't speak pure Hindi or pure English. They mix them. Models currently treat this as "bad grammar." They need to learn this as a valid dialect for reasoning.
• Capabilities: Authentic conversational bots for India; technical support systems that understand how Indian developers/engineers actually talk ("Code phat gaya production pe").
• Example Usage: A StackOverflow-style debugging thread written entirely in Hinglish.
• Schema Example:

  {
    "thread_id": "hinglish_debug_55",
    "context": "Python List Error",
    "conversation": [
      {"role": "user", "text": "Yaar, mera loop infinite chal raha hai, logic samajh nahi aa raha."},
      {"role": "expert", "text": "Check kar ki tune 'i' ko increment kiya hai ya nahi? Varna condition kabhi false nahi hogi."}
    ],
    "language_mix": ["Hindi", "English"]
  }
  

4. Indian Legal & Bureaucratic Reasoning

• Description: Analysis of Indian High Court judgments, land records, or tax laws in regional languages, annotated by legal experts.
• Why it's Critical: Law is jurisdiction-specific. US legal data is useless for Indian law. Models need to understand the specific logic of the Indian Penal Code (IPC) and Constitution in Marathi, Tamil, etc.
• Capabilities: AI legal assistants for Indian lawyers; automated document processing for Indian courts and government offices.
• Example Usage: A lawyer summarizing a Marathi High Court judgment, explaining the legal precedent used to reach the verdict.
• Schema Example:

  {
    "case_id": "BOM_HC_2024_33",
    "language": "Marathi",
    "verdict_summary": "...",
    "legal_reasoning_steps": [
      {"law_cited": "IPC Section 302", "interpretation": "This section applies because..."},
      {"precedent": "State vs. Shinde", "relevance": "Similar circumstances regarding evidence admissibility"}
    ]
  }
  

5. Indigenous Medical Knowledge (Ayurveda/Siddha)

• Description: Structured reasoning data from traditional Indian medical systems (Ayurveda, Siddha, Unani) in their native languages (Sanskrit, Tamil, Urdu).
• Why it's Critical: This is "novel knowledge" for the model. It adds a completely new scientific paradigm that doesn't exist in Western datasets.
• Capabilities: AI health assistants that can understand and respect traditional Indian medical practices alongside modern medicine.
• Example Usage: A Siddha practitioner diagnosing a patient based on "Naadi" (pulse) reading, explaining the logic in Tamil.
• Schema Example:

  {
    "diagnosis_id": "siddha_04",
    "symptoms": ["joint pain", "dry skin"],
    "system": "Siddha",
    "language": "Tamil",
    "reasoning": "Vata dosha imbalance detected due to...",
    "prescription": "Vathamadakki tailam",
    "contraindications": "Avoid cold foods"
  }
  

6. Hyper-Local Multimodal Understanding (Video + Causality)

• Description: Videos of Indian-specific physical tasks (e.g., weaving, pottery, street cooking, fixing a puncture) with expert narration explaining why actions are taken.
• Why it's Critical: For robotics and video models (Sora). A robot trained in a US kitchen fails in an Indian kitchen (different utensils, cooking methods). They need to learn "Indian physics."
• Capabilities: Service robots for India; advanced video understanding for safety and surveillance in chaotic Indian environments.
• Example Usage: A video of a mechanic fixing a Bajaj scooter engine, explaining in Hindi: "I am tightening this specific bolt first because vibration loosens it."
• Schema Example:

  {
    "video_id": "mechanic_puncture_repair",
    "language": "Bhojpuri",
    "timestamps": [
      {"start": "00:10", "end": "00:25", "action": "Locating leak", "reasoning": "Submerging tube in water to see bubbles"}
    ],
    "objects_involved": ["rubber_tube", "water_trough", "patch"]
  }
  

7. Dialect & Low-Resource Language Corpora

• Description: High-quality data for languages/dialects that have almost zero internet presence (Gondi, Santhali, Maithili, Tulu).
• Why it's Critical: To achieve "Universal Intelligence." We wants to serve every human. These "dark" data spots are their biggest missing piece.
• Capabilities: Preserving endangered languages; enabling digital access for tribal populations.
• Example Usage: Folk stories or daily conversations recorded in Santhali, transcribed and translated.
• Schema Example:

  {
    "language": "Santhali",
    "script": "Ol Chiki",
    "content_type": "Folk Tale",
    "audio_url": "s3://...",
    "transcript_original": "...",
    "transcript_latin": "...",
    "translation_english": "..."
  }
  

8. Native Indic Creative Writing & Nuance

• Description: Poems, satire, scripts, and lyrics written originally in regional languages, annotated for their "rasa" (emotional essence) and cultural metaphors.
• Why it's Critical: Current models sound "robotic" in Hindi. They translate English metaphors ("piece of cake") literally. They need to learn Indian idioms ("daal mein kuch kaala hai").
• Capabilities: High-quality creative writing assistants for Indian authors, scriptwriters, and marketers.
• Example Usage: A Hindi poem annotated line-by-line explaining the double meanings and cultural references.
• Schema Example:

  {
    "work_id": "poem_urdu_01",
    "language": "Urdu",
    "text": "...",
    "analysis": [
      {"line": 1, "metaphor": "Chand (Moon)", "cultural_meaning": "Beloved's face, not the satellite"},
      {"line": 2, "sentiment": "Melancholy (Udaasi)"}
    ]
  }