lumi-record-smh

Vietnamese smart-home voice recording dataset exported from lumi-record-voice-SMH.

Generated at: 2026-07-29T11:45:33

Configs

• mien-bac: Northern voices.
• mien-trung: Central voices.
• mien-nam: Southern voices.

All configs expose one split: test.

Columns

The dataset intentionally exposes only the columns needed for speech, speaker, and recording-context analysis.

Column	Type	Description
audio	Audio	Embedded WAV audio. In the Hugging Face Data Viewer this should render as an audio player.
text	string	Prompt text / transcription corresponding to the recording.
level	string	Prompt difficulty: easy, normal, or hard.
gender	string	Participant gender label: male or female.
participant_name	string	Participant name from manifest metadata.
region	string	Region folder/config: mien-bac, mien-trung, or mien-nam.
distance_label	string	Microphone distance label, for example 1m or 3m.
recording_mode	string	Recording type. SINGLE is one prompt per recording; CONVERSATION is a multi-turn prompt.
conversation_instance_id	string	Stable group ID shared by turns from the same multi-turn conversation. Empty for non-conversation samples when unavailable.
turn_index	int64	Turn order inside conversation_instance_id. Null for single-turn samples when unavailable.

For hard conversation samples, conversation_instance_id and turn_index are enough to reconstruct the recorded turns: group rows by conversation_instance_id, then sort each group by turn_index.

Usage

from datasets import load_dataset, Audio

ds = load_dataset("luvox-ai/lumi-record-smh", "mien-bac", split="test")
ds = ds.cast_column("audio", Audio())

Example: read a hard multi-turn conversation in order.

from datasets import load_dataset

ds = load_dataset("luvox-ai/lumi-record-smh", "mien-bac", split="test")
hard_turns = ds.filter(lambda row: row["level"] == "hard" and row["conversation_instance_id"])
conversation_id = hard_turns[0]["conversation_instance_id"]
conversation = hard_turns.filter(lambda row: row["conversation_instance_id"] == conversation_id)
conversation = conversation.sort("turn_index")

for row in conversation:
    print(row["turn_index"], row["text"])