
Closed
Posted
Paid on delivery
I need an end-to-end voice command system that runs entirely offline on a Rockchip board using the RKNN NPU. The solution must automatically identify the spoken language and then recognise a fixed set of voice commands without falling back to any cloud service. Target hardware The model will run on an embedded system powered by a Rockchip NPU (RK3588 / RK3568 class). Please optimise with RKNN-Toolkit and supply the final *.rknn file together with conversion scripts so I can reproduce the build from the original TensorFlow / PyTorch / ONNX model. Languages and use-case focus • Automatic language ID for: English, Mandarin Chinese, Cantonese, Thai, Korean, Japanese and Malay. • After language detection, the engine should map the utterance to a command ID (no large-vocabulary transcription needed). Deliverables 1. Trained multilingual model converted and benchmarked on the RKNN NPU. 2. Demo application (C/C++ or Python) that listens on the device microphone, prints the detected language and command, and triggers a GPIO or log action. 3. Clear documentation covering dataset choices, preprocessing, model architecture, quantisation steps, RKNN optimisation settings, and instructions for rebuilding or fine-tuning. Acceptance criteria • <150 ms end-to-end latency for a 1-second utterance on the target board. • ≥90 % command accuracy per language in clean conditions, with graceful handling of unknown input. • Runs fully offline with no external dependencies beyond RKNN runtime. If any additional libraries are required, keep them lightweight and open-source so the final firmware footprint remains small.
Project ID: 40770999
67 proposals
Remote project
Active 2 days ago
Set your budget and timeframe
Get paid for your work
Outline your proposal
It's free to sign up and bid on jobs
67 freelancers are bidding on average $485 USD for this job

Hi There, Hope you are well. I can develop an offline voice command system for Rockchip boards using the RKNN NPU, supporting multilingual language detection and fixed voice commands. I will deliver a trained model, demo application, and comprehensive documentation meeting all specified benchmarks and requirements, ensuring minimal external dependencies. Thanks in advance for the opportunity to working with you. Best Regards, Serazum
$300 USD in 1 day
7.6
7.6

The hardest part here is getting seven languages to share a single command classifier without the model collapsing into language confusion, especially when Mandarin and Cantonese share lexical overlap but differ phonetically. I would use a two-stage architecture: a lightweight language ID head trained on short audio segments, then route to language-specific command decoders that share a common feature extractor. That keeps the NPU footprint small and latency under your 150 ms target. I would start with a PyTorch model using a convolutional frontend for mel-spectrogram features, train the language ID classifier on Mozilla Common Voice and OpenSLR datasets, then fine-tune per-language command heads on your fixed vocabulary. Quantisation to INT8 happens in RKNN-Toolkit with per-channel calibration to preserve accuracy. The demo app will be in Python using pyaudio for the microphone stream, with GPIO triggers via the standard Linux sysfs interface. I will document every preprocessing step, the exact RKNN conversion flags, and calibration dataset so you can retrain or add commands later. What is the full list of voice commands you need recognised, and do you have any existing audio samples I should match for accent or recording conditions?
$750 USD in 21 days
7.5
7.5

Hello, I’ve worked on AI model deployment and hardware-focused optimisation, and I’m ready to get started. I’d use lightweight language identification and command classification models, converted through ONNX to RKNN for offline inference on RK3588/RK3568. Which exact board and RKNN runtime version are you using? Do you have labelled audio samples for all seven languages? I’d use INT8 quantisation and benchmark accuracy, memory, and latency to balance speed with reliability. I’ll build the pipeline, integrate microphone input and GPIO actions, then deliver the .rknn model, conversion scripts, benchmarks, and setup documentation. Best, Dev S.
$400 USD in 10 days
6.6
6.6

With a strong background in building and maintaining AI systems for production, I believe I am an excellent fit for your end-to-end multilingual ASR project. Being able to put AI where others can't, such as on edge devices like the Rockchip board you require, has been a core competency of mine. My expertise in firmware development and natural language processing paired with my proficiency in Python and Pytorch will enable me to deliver exactly what you need - a voice command system that runs offline and identifies spoken languages while recognizing fixed set of commands. I understand the ultra-offline nature of your project, and I am well-versed in designing lightweight and open-source solutions that keep the final firmware footprint minimal. My extensive experience with embedded systems and line-by-line optimization using tools like RKNN Toolkit will ultimately lead to a highly efficient model for your Rockchip NPU. Additionally, given my previous work in delivering across various hardware platforms including the ESP32/STM32 firmware, I am content handling low-level operations crucial for GPIO interaction or logging actions in your requested demo application.
$500 USD in 7 days
6.3
6.3

I am fully prepared to develop a multilingual voice command system for the Rockchip NPU (RK3588/RK3568) that operates 100% offline, with sub-150ms latency and ≥90% accuracy. My approach: First, I will train a language identification (Language ID) model and a fixed-command classifier for 7 languages using a lightweight convolutional/conformer architecture (in PyTorch). Subsequently, I will optimize the model via INT8/FP16 quantization using the RKNN-Toolkit2 to generate the final `.rknn` model. I will then build a low-memory-footprint demo application (in C++/Python) that captures microphone input via ALSA/PulseAudio on the edge hardware and triggers GPIO actions or logs based on the detected commands. Deliverables will include reproducible model conversion scripts, benchmarking data, and step-by-step tuning documentation. Key questions: 1. Is your target hardware the RK3588 or the RK3568, and which Linux OS (e.g., Debian/Ubuntu) is running on the board? 2. How many fixed commands need to be supported per language, and is the audio input single-channel (mono) at 16kHz? This solution will deliver ultra-fast, reliable, and production-ready performance on your embedded device—entirely without the need for the cloud.
$450 USD in 12 days
5.6
5.6

We have experience with AI models and embedded system development. We can assess the multilingual command recognition pipeline and optimize model conversion and inference for the Rockchip RKNN NPU. We will focus on offline operation and low latency with language detection and command classification. Could you share the exact board model and the number of commands required per language?
$500 USD in 7 days
5.7
5.7

Hi Ee, I will deliver a fully offline RKNN‑based multilingual ASR for RK3588/RK3568, with language ID for 7 languages, command mapping, demo app, *.rknn file and docs, under 150 ms latency. I can start today and provide a short test build
$500 USD in 3 days
5.3
5.3

Hi, I read your project description regarding an offline multilingual voice-command engine optimized for Rockchip RK3588/RK3568 NPU hardware. I understand you’re looking for a reliable professional to deliver the RKNN model, reproducible conversion scripts, embedded demo, GPIO or logging action, and documentation for English, Mandarin, Cantonese, Thai, Korean, Japanese, and Malay. I have successfully completed similar projects involving AI inference pipelines, model conversion, quantisation, Python/C++ integrations, and edge deployment, so I’m confident I can deliver a practical solution. My approach will be to first define the command dataset, preprocessing flow, language-ID strategy, and RKNN benchmark plan, followed by converting and optimizing the model for the target NPU and building the offline microphone demo. I’ll then validate latency, per-language accuracy, unknown-input handling, firmware dependencies, and the rebuild documentation against your acceptance criteria. I prioritize compact, reproducible implementations and clear technical communication, especially for embedded deployments where the 150 ms target matters. Would you be open to a quick chat about the exact command set and target board configuration? Regards, Ravisher
$750 USD in 7 days
4.9
4.9

Hi, I have 10+ years of experience building end-to-end offline speech pipelines with Python and hardware-oriented optimizations. I will deliver an offline RKNN-compatible multilingual language ID plus fixed command mapping workflow for the RK3588/RK3568 class, including conversion scripts to reproduce the .rknn from your original TensorFlow/PyTorch/ONNX models. I will provide a demo app in Python that prints detected language and command and triggers a GPIO/log action with clear RKNN Toolkit quantization and benchmark documentation.
$300 USD in 5 days
5.0
5.0

Your RKNN pipeline will fail if you quantize the multilingual ASR model without per-language calibration datasets - this causes accuracy to drop below 70% for tonal languages like Cantonese and Thai when you compress to INT8. Quick questions - what's your target power budget for continuous inference, and have you already collected command samples across all 7 languages or do you need synthetic augmentation? Here is the architectural approach: - RKNN-TOOLKIT QUANTIZATION: Build language-specific calibration sets (500+ samples per language) and use hybrid quantization to preserve accuracy above 90% while hitting your 150ms latency target on RK3576. - MULTILINGUAL ASR PIPELINE: Deploy a two-stage model - lightweight language ID classifier (15ms) feeding into language-specific command recognizers, all compiled to .rknn with shared feature extraction to minimize SRAM footprint. - EMBEDDED DEMO APPLICATION: Deliver C++ inference wrapper with ring-buffer audio capture, VAD preprocessing, and GPIO trigger logic, plus Python conversion scripts that reproduce the ONNX-to-RKNN build from your training checkpoints. I've deployed similar edge ASR systems for 2 IoT manufacturers running on Rockchip NPUs with <100ms latency. Let's schedule a 20-minute technical call to review your command vocabulary and calibration data before I finalize the architecture.
$450 USD in 10 days
5.5
5.5

Hello, I understand you need a fully offline multilingual voice-command system optimized for Rockchip RK3576/RK3588-class hardware, using the RKNN NPU without cloud services. I would structure the solution around lightweight language identification and command classification for English, Mandarin, Cantonese, Thai, Korean, Japanese, and Malay. My approach would cover dataset and preprocessing selection, model evaluation, ONNX/TensorFlow/PyTorch conversion, RKNN optimization and quantization, followed by on-device benchmarking. I would build a C/C++ or Python demo that captures microphone audio, identifies the language, returns a command ID, and triggers a GPIO or log action, including handling unknown commands safely. I would also provide reproducible conversion scripts, configuration details, and clear documentation for rebuilding and fine-tuning. Latency below 150 ms and accuracy of at least 90% would be treated as measurable acceptance targets, validated on your actual board rather than assumed in advance. Looking forward to work with you. Thanks
$320 USD in 3 days
5.0
5.0

Hi there, You need a fully offline multilingual voice-command engine on RK3588/RK3568 that uses the RKNN NPU, identifies seven languages, recognizes fixed commands, and stays below 150 ms latency without cloud dependencies. I have strong hands-on experience with Python, embedded inference, and model deployment, and have solved similar problems involving ONNX/TensorFlow conversion, quantisation, and low-latency audio pipelines. My approach would be to: * Review the source model, command set, audio format, board environment, and current implementation. * Identify the root cause of latency or accuracy limitations and determine what can be reused versus what needs to be changed. * Implement a clean, maintainable solution using Python, RKNN-Toolkit, RKNN Runtime, and a lightweight microphone/GPIO integration. * Test language identification, unknown-input rejection, per-language accuracy, quantisation behavior, latency, reliability, and offline operation. * Deploy or integrate the fix without disrupting the existing firmware, while providing the .rknn artifact, conversion scripts, benchmarks, and rebuild documentation. A cascaded language-ID and command classifier can reduce model size and inference cost, but confidence thresholds must be calibrated jointly to prevent incorrect commands when speech is ambiguous. I prefer fixing the actual root cause rather than applying temporary patches, and I will keep the implementation simple, stable, and maintainable. I can start by rev
$650 USD in 21 days
4.6
4.6

Hi, I am an embedded systems engineer with over 16 years of experience. I have extensive experience with embedded Linux, on-device machine learning and real-time hardware integration. I can build an offline system that identifies the spoken language and maps supported phrases to command IDs across all seven languages. My approach would use compact language identification and command recognition models, optimised for RKNN, with per-language accuracy checks and rejection of unknown input. I will provide the .rknn model, reproducible conversion scripts, a microphone demo with GPIO or log actions, and documentation covering training data, preprocessing, quantisation and rebuilding. I can also support firmware integration and the practical steps needed to turn the prototype into a reliable product. Could you confirm the exact board? The title mentions RK3576, while the description lists RK3588/RK3568. Do you have the command list and labelled recordings for each language? Also, is the 150 ms latency measured after the one-second utterance finishes? My 14-day estimate assumes a defined command set, suitable training data and access to the target hardware for benchmarking against your acceptance criteria. Please contact me to discuss details.
$750 USD in 14 days
4.8
4.8

Your requirement for a fully offline multilingual voice command engine on Rockchip NPU hinges on two critical bottlenecks: keeping feature extraction overhead minimal to guarantee sub-150ms latency, and preventing precision loss during RKNN INT8 quantization. I will implement a lightweight dual-head acoustic architecture (compact CNN/Conformer backbone) trained in PyTorch, outputting simultaneous Language ID across the 7 languages and fixed command classification with an unknown/rejection class. Implementation roadmap: ✅ Feature pipeline and model: 80-channel log-mel spectrogram preprocessor paired with an efficient acoustic model exporting cleanly to ONNX with dynamic batching. ✅ RKNN conversion and INT8 quantization: RKNN-Toolkit2 pipeline with balanced calibration datasets across English, Mandarin, Cantonese, Thai, Korean, Japanese, and Malay to maintain 90%+ accuracy. ✅ Edge demo app: Low-latency Python or C++ runtime utilizing librknnrt, ALSA ring-buffered mic stream, sliding-window inference, and libgpiod hardware triggering. ✅ Deliverables and docs: PyTorch/ONNX source, conversion scripts, final .rknn models, and step-by-step fine-tuning documentation.
$425 USD in 1 day
4.3
4.3

Hello. I've successfully tackled complex embedded AI projects involving offline speech recognition, multilingual command classification, ONNX-to-RKNN conversion, and Rockchip NPU optimization, turning challenging hardware constraints into fast, reliable, production-ready solutions. Two key questions: 1. Which exact Rockchip board will be used for final deployment, RK3576, RK3588, or RK3568, and what RKNN runtime version is installed? 2. Do you already have labeled voice-command datasets for all seven languages, including Cantonese and Thai, or should dataset preparation and augmentation be included? Your answers will greatly help me understand the project and implement an efficient pipeline covering automatic language identification, lightweight command classification, INT8 quantization, RKNN conversion, on-device benchmarking, GPIO integration, and reproducible documentation. I can start immediately and focus on what matters most: a genuinely offline, compact, reproducible solution engineered to achieve your 150 ms latency and 90% per-language accuracy targets, with proper unknown-command rejection and no hidden cloud dependencies. Let's discuss this project further Best regards, Jeff Chong.
$500 USD in 7 days
4.1
4.1

Greetings, I worked on an MStar 7T00/7T01 based KTV system where I handled embedded audio processing, and I’m also familiar with TTS/ASR systems and SDKs such as Cerence, so this kind of offline voice pipeline is familiar to me. For this project I would keep the model small and command-focused instead of doing full speech transcription, with language ID first and then multilingual command classification optimized for the Rockchip NPU. The main challenge will be keeping accuracy after INT8 quantisation while also meeting the 150 ms target, so I would first validate the model operators, RKNN conversion and real NPU latency before spending time on the full dataset. I can provide a first RKNN feasibility demo in about 5–7 days, including language/command inference and basic microphone input, then continue with training, unknown-command handling and device integration. Please let me know the exact Rockchip board and approximately how many commands you need per language. Sincerely, Oleksandr
$500 USD in 7 days
3.8
3.8

Hello, client, I can help build an entirely offline multilingual voice-command system for RK3588/RK3568, combining automatic language identification across all seven target languages with fast, fixed-command classification on the Rockchip NPU. I’ll evaluate a compact model, optimize and quantize it with RKNN-Toolkit, and deliver the `.rknn` model, reproducible conversion scripts, and a C/C++ or Python demo with microphone input, unknown-command rejection, and GPIO/logging actions. To meet your firmware-footprint requirement, I’ll minimize model size, runtime dependencies, and memory usage; prefer lightweight open-source audio preprocessing and RKNN Runtime; and exclude cloud services, large transcription engines, and unnecessary libraries. I’ll document the final model size, runtime footprint, quantization settings, and measured latency and accuracy on the target board. I’ll also validate the complete pipeline against your acceptance criteria, including the 150 ms latency target and 90% per-language command accuracy, reporting any hardware-dependent limitations transparently. Could you confirm the exact board model and available RAM/flash storage so I can optimize the deployment accordingly? Best regards, Darial
$500 USD in 7 days
3.9
3.9

==== Hi - Truong here ==== "OFFLINE MULTILINGUAL VOICE COMMANDS ON ROCKCHIP RKNN" - you need seven-language command recognition on the RK3576-class device without cloud calls, while meeting strict latency and accuracy targets. I would separate language identification from fixed-command classification, then benchmark quantized RKNN models on the exact board and runtime. The key risk is meeting the 150 ms target across all seven languages, so I would measure the full microphone-to-command path early and test unknown speech to prevent false triggers. I would provide conversion scripts, a demo with GPIO or log output, and reproducible build and fine-tuning documentation. Which exact board and RKNN runtime version will you use, and do you already have command recordings for all seven languages? Looking forward to work with you.
$323 USD in 7 days
3.9
3.9

Nice to meet you , It is a pleasure to communicate with you. My name is Anthony Muñoz, I am the lead engineer for DSPro IT agency and I would like to offer you my professional services. I have more than 10 years of working as a Backend and Software developer, I have successfully completed numerous jobs similar to yours therefore, and after carefully reading the requirements of your project, I consider this job to be suitable to my area of knowledge and skills. I would love to work together to make this project a reality. I greatly appreciate the time provided and I remain pending for any questions or comments. Feel free to contact me. Greetings
$462 USD in 7 days
3.8
3.8

Hello, Your project requires a fully offline multilingual voice-command pipeline on Rockchip RK3588 or RK3568, combining language identification, command classification, low latency, and reproducible RKNN deployment without cloud dependencies. The main challenge is meeting the accuracy and sub-150 ms target within a compact embedded runtime. I would structure the solution as a lightweight two-stage model, with shared audio preprocessing, language detection, command classification, unknown-input handling, INT8 quantisation, and RKNN NPU benchmarking. I will provide the converted .rknn model, conversion scripts, and a C++ or Python demo that connects microphone input to GPIO or logging. Similar work includes Covertly AI, where I built AI-powered platforms with Node.js APIs and model integrations. My experience with Python, ONNX workflows, embedded-oriented optimisation, and offline inference supports a practical implementation focused on measurable device performance. I will deliver a clean first working milestone within 7 days, including model conversion, device benchmarking, demo integration, and concise documentation for rebuilding and fine-tuning. Best regards, Bohdan
$700 USD in 7 days
3.9
3.9

Penang , Butterworth, Malaysia
Payment method verified
Member since Dec 13, 2017
$250-750 USD
$30-250 USD
$250-750 USD
$250-750 USD
$30-250 USD
£250-750 GBP
$15-25 USD / hour
₹100-400 INR / hour
$30-250 USD
$25-50 USD / hour
$2-8 USD / hour
$250-750 USD
$10-30 USD
$8-15 USD / hour
$250-750 USD
$30-499 USD
$30-250 USD
$15-25 USD / hour
₹12500-37500 INR
₹600-1500 INR
$750-1500 USD
₹750-1250 INR / hour
₹37500-75000 INR
₹12500-37500 INR
₹1500-12500 INR