NeuDecide: Turn speech into actions on-device.

NeuDecide: Turn speech into actions on-device.

NeuDecide: Turn speech into actions on-device.

Build private voice control with a 43 MB decision model that turns spoken requests into tool calls. Run locally on a single CPU thread, without a transcription step.

Build private voice control with a 43 MB decision model that turns spoken requests into tool calls. Run locally on a single CPU thread, without a transcription step.

Build private voice control with a 43 MB decision model that turns spoken requests into tool calls. Run locally on a single CPU thread, without a transcription step.

Tell your devicewhat to do.RecordTRY SAYING“Vacuum the kitchen.”“Mop the bathroom.”“Return to yourdock.”“How much battery isleft?”“How full is thedustbin?”ModelTool selection{ }OutputLivingKitchenBedroomHallBathOfficeDOCK72%30%Robot vacuumiLEFTRIGHTRobot armi05:00SmartwatchiLampi
Robot vacuumRobot armSmartwatchLampTell your device what to do.RecordModelTool selection{ }STATUSPick a device, then press Record.The tool call the model returns shows up here.LivingKitchenBedroomHallBathOfficeDOCK72%30%Robot vacuumiTRY SAYING“Vacuum the kitchen.”“Mop the bathroom.”“Return to your dock.”“How much battery is left?”“How full is the dustbin?”

Interact with our demo below. Choose a device. Tell it what to do.

Interact with our demo below. Choose a device. Tell it what to do.

Select a device then record your voice. NeuDecide turns your words into actions.

Select a device then record your voice. NeuDecide turns your words into actions.

Select a device then record your voice. NeuDecide turns your words into actions.

[

HOW IT WORKS

]

Speech to tool calls, trained as one system.

Speech to tool calls, trained as one system.

Speech to tool calls, trained as one system.

0.1

Audio Encoder

Audio encoder turning speech into compact audio representations

Listen to the request.

Converts speech into compact audio representations, without transcription.

0.1

Audio Encoder

Audio encoder turning speech into compact audio representations

Listen to the request.

Converts speech into compact audio representations, without transcription.

0.2

Tool Encoder

Tool encoder combining speech with your tool definitions

Connect speech to your tools.

Combines audio representations with your tool definitions, preparing the context for decoding.

0.2

Tool Encoder

Tool encoder combining speech with your tool definitions

Connect speech to your tools.

Combines audio representations with your tool definitions, preparing the context for decoding.

0.3

Tool Call Decoder

Tool call decoder generating the final tool call

Build the tool call.

Generates the call one token at a time, reusing cached attention state to keep each step efficient.

0.3

Tool Call Decoder

Tool call decoder generating the final tool call

Build the tool call.

Generates the call one token at a time, reusing cached attention state to keep each step efficient.

[

KEY FEATURES

]

Built for voice control on-device.

Built for voice control on-device.

Built for voice control on-device.

Define your tools and turn spoken requests into structured calls, with a compact model that runs locally on CPU.

Define your tools and turn spoken requests into structured calls, with a compact model that runs locally on CPU.

Define your tools and turn spoken requests into structured calls, with a compact model that runs locally on CPU.

Feature

Feature

Feature

Specification

Specification

Specification

Model width

Model width

Model width

512

512

512

Decoder layers

Decoder layers

Decoder layers

8

8

8

Attention

Attention

Attention

Grouped-query, 4 key/value heads of 64 dimensions

Grouped-query, 4 key/value heads of 64 dimensions

Grouped-query, 4 key/value heads of 64 dimensions

Vocabulary

Vocabulary

Vocabulary

8,192 SentencePiece BPE tokens

8,192 SentencePiece BPE tokens

8,192 SentencePiece BPE tokens

Sample rate

Sample rate

Sample rate

16 kHz

16 kHz

16 kHz

Limits

Limits

Limits

30 s of audio, 1,536 tool tokens, 128 output tokens

30 s of audio, 1,536 tool tokens, 128 output tokens

30 s of audio, 1,536 tool tokens, 128 output tokens

Device

Device

Device

Time to call

Time to call

Time to call

Load time

Load time

Load time

MacBook Pro M3

MacBook Pro M3

MacBook Pro M3

46 ms

46 ms

46 ms

159 ms

159 ms

159 ms

Samsung S24+

Samsung S24+

Samsung S24+

82 ms

82 ms

82 ms

267 ms

267 ms

267 ms

Raspberry Pi 5

Raspberry Pi 5

Raspberry Pi 5

206 ms

206 ms

206 ms

499 ms

499 ms

499 ms

  • Time to call stayed below 210 ms across all three devices.

  • Loading took less than half a second.

  • Peak RAM stayed below 175 MB: ranging from 146 MB to 174 MB across the tested devices.

  • Time to call stayed below 210 ms across all three devices.

  • Loading took less than half a second.

  • Peak RAM stayed below 175 MB: ranging from 146 MB to 174 MB across the tested devices.

  • Time to call stayed below 210 ms across all three devices.

  • Loading took less than half a second.

  • Peak RAM stayed below 175 MB: ranging from 146 MB to 174 MB across the tested devices.

Download NeuDecide.

Download NeuDecide.

Download NeuDecide.

Get our 43 MB model and run speech-to-tool calling locally on your own hardware.

Get our 43 MB model and run speech-to-tool calling locally on your own hardware.

Build with NeuDecide.

Build with NeuDecide.

Build with NeuDecide.

Get the Python package, define your tools, and turn spoken requests into structured calls.

Get the Python package, define your tools, and turn spoken requests into structured calls.

Bring your products to life.

Bring your products
to life.

Talk to us about integrating NeuDecide into your robots, devices, or applications.

Talk to us about integrating NeuDecide into your robots, devices, or applications.