somu9/hindi-speech-instruct

Hindi Conversational Speech Dataset

Multi-turn Hindi conversational dataset for training speech language models. Each conversation has user audio turns paired with assistant text responses.

Dataset Statistics

Metric	Value
Total conversations	10
Total user turns (audio files)	25
Total assistant turns	25
Min turns per conversation	1
Max turns per conversation	4
Avg user text length	32.4 chars
Avg assistant text length	94.0 chars

Turn Distribution

Turns	Count	Percentage
1	2	20.0%
2	3	30.0%
3	3	30.0%
4	2	20.0%

Dataset Structure

dataset/
├── conversations.jsonl
└── audio/
    ├── conv_001_turn_001.mp3
    ├── conv_002_turn_001.mp3
    └── ...

conversations.jsonl Format

Each line is a JSON object representing one conversation:

{
  "conversation_id": "conv_001",
  "turn_count": 2,
  "turns": [
    {"role": "user", "audio": "audio/conv_001_turn_001.mp3", "text": "..."},
    {"role": "assistant", "text": "..."},
    {"role": "user", "audio": "audio/conv_001_turn_002.mp3", "text": "..."},
    {"role": "assistant", "text": "..."}
  ]
}

• User turns have audio (path to mp3) and text (transcript)
• Assistant turns have only text (the response)
• Turns always alternate: user → assistant → user → assistant
• turn_count = number of user turns in the conversation

Audio Naming Convention

{conversation_id}_turn_{NNN}.mp3 — e.g., conv_001_turn_001.mp3

Sample Conversation

User 🎤: आज दिल्ली में मौसम कैसा रहेगा?

Assistant: आज दिल्ली में काफ़ी गर्मी रहेगी, दोपहर में तापमान चालीस के पार जा सकता है, शाम को थोड़ी राहत मिलेगी।

Intended Use

Training speech language models that take audio input and generate text responses (replacing the ASR → LLM pipeline with a single model).

Language

All user inputs and assistant responses are in Hindi (हिंदी).

---

Last updated: 2026-06-17 18:51 UTC