lokeshe09/bharat-voice-conversations

Bharat Voice Conversations

Multi-turn, code-mixed, tool-calling conversations for fine-tuning the LLM inside Indian-language voice agents (call centres, IVR replacements, helplines).

Most open Indic data is read speech, translated instructions or single-turn Q&A. Real voice agents need something else: short spoken replies, callers who mix English into their language (Hinglish, Tenglish), misheard digits from speech recognition, API/tool calls in the middle of the conversation, and safe handling of emergencies and fraud. This dataset targets exactly that gap.

Config	Language	Conversations	Train / Test
telugu	Telugu (తెలుగు), Telugu script + Romanized	2,500	2,294 / 206

More Indian languages will be added as new configs in this same repo.

Quick start

from datasets import load_dataset

ds = load_dataset("lokeshe09/bharat-voice-conversations", "telugu")

# filter by any metadata field
banking   = ds["train"].filter(lambda r: r["sector"] == "bfsi")
romanized = ds["train"].filter(lambda r: r["style"] == "romanized_codemixed")
handoffs  = ds["train"].filter(lambda r: r["outcome"] == "escalated_to_human")

Train with tools (OpenAI-style chat format, works with most chat templates):

import json

def to_chat(r):
    return {"messages": r["messages"], "tools": json.loads(r["tools"])}

What's inside

• 103 domains, 561 subdomains, 17 sectors. Covers banking/UPI, loans and collections, insurance, e-commerce, telecom, utilities, healthcare, government services and welfare schemes, agriculture, travel and railways, education, HR/IT helpdesks, legal intake, emergency, women's and child helplines, and more. Every subdomain has at least 4 conversations.
• Two styles, 50/50.
  • native_codemixed: Telugu script with natural English words, e.g. "మీ recharge success అయింది".
  • romanized_codemixed: Latin-script Tenglish, e.g. "Mee refund process chesanu andi".
• Six call types. inbound support (1,966), outbound reminders (223), verification (142), sales (95), surveys (51), collections (23).
• Realistic outcomes. resolved (1,633), callback scheduled (338), escalated to a human (285), declined (131), unresolved (113).
• Speech-recognition noise. 300 conversations contain raw-ASR-style caller text, with no punctuation and missing or misheard digits that the agent asks the caller to confirm.
• Tool calling. 5,470 tool calls in total. Each conversation declares its own tools (JSON schema), and every call is followed by its result.
• Varied callers. Regions span Telangana, Coastal Andhra and Rayalaseema. Callers include elderly people, farmers, students, small-business owners, NRI families, angry repeat callers and callers wary of fraud.
• Length. On average 7.1 caller turns per conversation, and 545 conversations have 10 or more.

Agent behaviour the data teaches

• Replies are one or two short spoken sentences, with numbers, amounts and dates written as words so text-to-speech reads them correctly.
• The agent reads back IDs and digits to confirm them, and confirms before taking any action.
• It never asks for OTP, PIN, CVV or passwords, and warns callers about scams.
• It states facts only from tool results or the system prompt, and says "I don't have that information" instead of inventing.
• Safety-critical domains (medical, mental health, women's and child safety, disasters):
  • it screens for danger first and never diagnoses or prescribes;
  • it routes callers to 112, 108, 181, 1098, 1930, 14416 and other helplines;
  • it hands off to humans with a summary.
• Collections calls are polite and factual, with no threats.

Data fields

Field	Description
id	Unique ID, e.g. te_000001
language, language_code	telugu, te
script	native or latin
style	native_codemixed or romanized_codemixed
sector, domain, subdomain	Three-level taxonomy (17 / 103 / 561)
call_type	inbound_support, outbound_reminder, outbound_sales, outbound_collections, survey_feedback, verification
region	City, town or district of the caller
user_persona, user_emotion	Short caller description and emotional state
outcome	resolved, callback_scheduled, escalated_to_human, declined, unresolved
asr_noise	true if the caller text simulates raw speech-recognition output
num_turns, num_tool_calls	Caller turns and tool calls in the conversation
quality_tier	auto_checked (see Limitations)
tools	JSON string: list of tool definitions (OpenAI function format)
messages	List of {role, content, tool_calls, tool_call_id, name} with roles system, user, assistant, tool

Example (abridged)

{"role": "user", "content": "హలో, నేను మూడు వందల తొంభై తొమ్మిది రూపాయల recharge చేశాను, డబ్బులు కట్ అయ్యాయి కానీ recharge అవ్వలేదు."}
{"role": "assistant", "content": "క్షమించండి అండి. మీరు recharge చేసిన మొబైల్ నంబర్ చెప్తారా?"}
{"role": "user", "content": "9848206715"}
{"role": "assistant", "content": "ఒక్క నిమిషం అండి, చెక్ చేస్తున్నాను.", "tool_calls": [{"id": "call_1", "type": "function", "function": {"name": "get_recharge_history", "arguments": "{\"mobile_number\": \"9848206715\"}"}}]}
{"role": "tool", "tool_call_id": "call_1", "name": "get_recharge_history", "content": "{\"transactions\": [{\"transaction_id\": \"TXN78451236\", \"amount\": 399, \"status\": \"PENDING\"}]}"}
{"role": "assistant", "content": "మూడు వందల తొంభై తొమ్మిది రూపాయల recharge pending లో ఉంది అండి. దాన్ని ఇప్పుడే మళ్ళీ process చేయమంటారా?"}

How it was made

• Synthetic and AI-written. Every conversation was written by an AI model (Anthropic's Claude) working conversation by conversation against a fixed taxonomy and slot plan. They are not transcripts of real calls.
• No real people's data. All names, phone numbers, account numbers and IDs are fictional. Agent-side companies are fictional (e.g. "Pragati Bank", "Bazaarly", "Sravya Telecom"). Real public bodies, schemes and helplines (e.g. RBI Ombudsman, PM-KISAN, Ayushman Vay Vandana, 1930, 108) are named where relevant.
• Automatic checks on every conversation:
  • taxonomy match;
  • tool-call/result pairing;
  • tool-argument types, enums and required fields;
  • correct script per style (no Telugu characters in Romanized rows);
  • no duplicates;
  • continuous IDs.
• Dates. The calendar context is fixed at Saturday, 17 October 2026, so relative dates ("tomorrow", "Monday") are consistent.

Limitations

• No native-speaker review yet. All rows are quality_tier: auto_checked. Expect occasional unnatural phrasing; a native review pass is planned.
• Policy facts can go out of date. Real-world rules mentioned in system prompts (tax thresholds, scheme amounts, RBI timelines) were correct to the best of our knowledge in 2026 but may change. Some tool values, such as prices and support prices, are illustrative only.
• Uneven call types. Inbound support dominates; collections calls are relatively few.
• Text only. There is no audio. ASR noise is simulated in text.
• Not for real advice. Do not use the data as a source of legal, medical or financial guidance.

License

CC BY 4.0. You may use, share and adapt the data, including commercially, with attribution.

Citation

@dataset{bharat_voice_conversations_2026,
  title  = {Bharat Voice Conversations: Code-mixed, tool-calling voice-agent dialogues for Indian languages},
  year   = {2026},
  url    = {https://huggingface.co/datasets/lokeshe09/bharat-voice-conversations}
}