Skip to content
HOME / PRODUCTS / VOXEDGE
VOXEDGE / ON-DEVICE MODELS

The engine
behind every voice.

Listening, translating and speaking, all on the device: VAD, ASR, MT and TTS, on chips from MediaTek Helio G81 and above.

01 / FOUR STRENGTHS

What the engine
gives you.

Select the acoustic and translation capabilities your experience needs, then evaluate quality and resource use on your target hardware.

01 / NO CLOUD BILL
$0per callcloud cost

Nothing per call.

Recognition, translation and speech run on the device, so using them never adds to a cloud bill.

NETWORK / ROUND TRIPS
0 network round trips
No per-call cloud cost. SDK licensing applies.
02 / FASTER THAN A BLINK
~70mson-device speech latency

Keep up with the conversation.

Speech turns into text as fast as people talk, with no waiting on a network.

ASR / RECOGNITION QUALITY
8.02% English word error rate
8.91% Chinese character error rate
Internal noisy-speech evaluation at ≥10 dB signal-to-noise ratio. Lower is better. Proprietary always-on ASR model.Proprietary always-on ASR: 65M parameters. VoxSense reference package approximately 80 MiB. Footprint varies by model and configuration.
03 / BARELY ANY BATTERY
180mAh / 12hestimated incremental use

Built to stay aware.

Low-power audio capture and local transcription, optimized together across hardware and software.

POWER / BATTERY IMPACT
12 h estimated consumption
Estimated 12-hour incremental use: 180 mAh, based on 15 mAh per hour. Reference setup: MTK9600, continuous speech playback, transcription once per minute.
04 / MULTILINGUAL REACH
103languages & writing variants

Speak to more markets.

Multilingual translation coverage, with speech capabilities focused on underserved local languages.

SPEECH / LANGUAGE COVERAGE
36 On-device ASR languages
45 On-device TTS languages
103 refers to on-device translation languages and writing variants. ASR and TTS coverage are listed separately; availability varies by configuration.
TRANSLATION / BENCHMARK RESULTS
+17.5% relative BLEU improvement vs. Hy-MT2*
20 / 21 mainstream directions won vs. Google Translate offline
*548 MiB balanced model vs. Hy-MT2-1.8B (1.25-bit): internal evaluation of all 32 shared English-to-target directions. BLEU relative improvement, not accuracy. Google result: separate 239 MiB compact model, internal evaluation limited to 21 mainstream directions.Compact model: 239 MiB package, approximately 320 MiB peak memory; 121 tokens/s on Helio G85.

INTERNAL TEST RESULTS · CONFIGURATION-SPECIFIC METRICS

02 / INTEGRATION

Four steps.
Simple to integrate.

01

Create an application

Enter your app’s platform and package name, then choose the capabilities and languages to try.

02

Approval and NDA

Wait for approval and sign the NDA.

03

Download the SDK

Build and download the SDK for your platform. It includes every capability in your license.

04

Initialize with your key

Pass your application key at start-up. The device activates against the license on first use.

Get started ↗
03 / QUESTIONS

Before you build.

How low-end can the device be?

VoxEdge is designed for broad chipset compatibility, so it can reach entry-level platforms. It runs on MediaTek Helio G81 and above. Latency, power and memory vary by chipset and configuration, so validate on your target hardware.

Do all capabilities run on the device?

VoxEdge provides device-side acoustic and translation models. Your selected model packages, application and audio pipeline determine the final configuration.

Are translation and speech language counts the same?

No. Translation covers 103 languages and writing variants; ASR and TTS cover 36 and 45 languages respectively. Confirm the intersection required for your selected speech translation pairs.

Do the measured results apply to every device?

Results are specific to their test configurations. Validate your chipset, language pairs and audio workflow before committing to latency, power and memory targets.

BUILD WITH VOXNEXUS

Your next product.
[ on-device AI ]

CONTACT US ↗