nandinireddy123/audio-function-calling

Audio Function Calling Dataset

A synthetic multi-turn conversation dataset for audio-based tool/function calling. Each sample contains a system prompt with tool definitions, alternating user and assistant turns, where user turns are designed for audio (natural spoken language with filler words) and assistant turns may include tool calls.

Note: This is the first batch (137 audio samples, 153 text-only samples). We are actively improving the generation pipeline and will be adding more data soon.

Structure

Subset	Samples	Description
text_only/	153	Text-only conversations (no audio)
audio/	137	Conversations with TTS-generated audio for user turns

Schema

text_only/conversations.parquet

Column	Type	Description
sample_id	string	Unique sample identifier
persona	string	Description of the user persona (from liquidrandom)
tools_json	string	JSON array of tool definitions in OpenAI function calling format
conversation	string	JSON array of conversation turns (role, content, optional tool_calls/tool_call_id)
num_turns	int32	Number of turns in the conversation
num_tool_calls	int32	Number of tool calls made by the assistant

audio/conversations.parquet

All columns from text_only plus:

Column	Type	Description
voice	string	TTS voice used for audio generation
user_audio	list<binary>	WAV audio bytes for each user turn

Generation Process

1. Seeding: Each sample is seeded with a random persona and 1-3 tool groups from liquidrandom
2. Conversation generation: An LLM generates multi-turn conversations (2-5 turns) with tool calls, designed for spoken language
3. Quality validation: A second LLM pass validates tool call correctness (no hallucinated arguments, matching tool definitions)
4. Audio synthesis: User turns are converted to audio using TTS

Usage

import pyarrow.parquet as pq

# Load text-only conversations
text = pq.read_table("text_only/conversations.parquet")

# Load audio conversations
audio = pq.read_table("audio/conversations.parquet")

# Access a sample
import json
row = audio.to_pydict()
conversation = json.loads(row["conversation"][0])
tools = json.loads(row["tools_json"][0])
user_audio_blobs = row["user_audio"][0]  # list of WAV bytes

License

MIT