cagataydev/vlm-voice-audio

🎙️ VLM Robotics Voice Commands (Audio)

Natural speech commands for Vision-Language-Model robot control.

This dataset contains 5,000 audio recordings of human voice commands for controlling robots — covering pick & place, navigation, manipulation, observation, multi-step tasks, spatial commands, safety, household chores, and conversational feedback.

🎯 Purpose

Training omni-modal VLMs that understand spoken robot commands. The audio captures natural speech patterns including:

• Polite forms: "Could you pick up the red cup?"
• Casual: "Hey, grab that"
• Urgent: "Stop immediately!"
• Multi-step: "Open the drawer, take out the spoon, and close it"
• Context-dependent: "The one I pointed at" (requires vision)

📊 Statistics

Metric	Value
Total Examples	5,000
Total Audio	3.3 hours
Avg Duration	2.4s
Sample Rate	44.1 kHz
Format	WAV (16-bit)
Language	English

Category Distribution

Category	Count	%
pick_place	1,214	24.3%
manipulation	770	15.4%
multistep	612	12.2%
navigation	593	11.9%
observation	523	10.5%
spatial	406	8.1%
household	321	6.4%
safety	224	4.5%
conversational	182	3.6%
context_rich	155	3.1%

Voice Distribution

Voice	Count	%
NATM1	2,519	50.4%
NATF0	523	10.5%
NATM0	484	9.7%
NATF1	398	8.0%
NATM2	322	6.4%
NATF2	292	5.8%
NATF3	235	4.7%
NATM3	227	4.5%

Difficulty Distribution

Difficulty	Count
Easy	406
Medium	3,827
Hard	767

🏗️ Schema

Column	Type	Description
audio	Audio	WAV audio at 44.1 kHz
text	string	Transcript of the spoken command
voice	string	Speaker voice ID (NATM0-3, NATF0-3)
category	string	Command category (pick_place, navigation, etc.)
difficulty	string	easy / medium / hard
duration_seconds	float	Audio duration in seconds
id	string	Unique example ID

🗣️ Voice Profiles

Voice	Description
NATM1	Natural male — calm, clear (primary)
NATM0	Natural male — low-pitched, measured
NATM2	Natural male — warm, friendly
NATM3	Natural male — deep, authoritative
NATF0	Natural female — clear, professional
NATF1	Natural female — energetic, upbeat
NATF2	Natural female — soft, thoughtful
NATF3	Natural female — bright, enthusiastic

📝 Example Commands

Pick & Place:

• "Pick up the red cube on the table"
• "Grab the bottle from the shelf and bring it here"
• "Could you move the blue box to the left side?"

Navigation:

• "Go to the kitchen"
• "Move forward two meters"
• "Turn left and face the door"

Manipulation:

• "Open the drawer slowly"
• "Pour water into the cup"
• "Fold the towel and put it away"

Multi-step:

• "First go to the counter, pick up the mug, then bring it to me"
• "Clear the table and organize everything in the bin"

Safety:

• "Stop immediately!"
• "Be careful, that's fragile"
• "Slow down"

🔧 Usage

from datasets import load_dataset

ds = load_dataset("cagataydev/vlm-voice-audio", split="train")

# Listen to a sample
example = ds[0]
print(f"Text: {example['text']}")
print(f"Category: {example['category']}")
# example["audio"] contains the waveform array + sampling_rate

📦 Related Datasets

Dataset	Description
cagataydev/vlm-voice-commands	Text-only version (50K commands)
cagataydev/omni-voice-training	General Q voice training data
cagataydev/q-omni-data-soup	Multi-modal training dataset

🏭 Generation

Generated using Parler-TTS Mini v1.1 on NVIDIA L40S (46GB VRAM). Speaker characteristics controlled via natural language descriptions. Source text from curated template-based generation covering 10 robotics command categories.

Built with DevDuck 🦆

---

Generated on 2026-03-23