How Audio-Visual Speech Recognition Works Under the Hood
TL;DR
A complete, up-to-date breakdown of under the hood for developers and founders. It covers the core ideas, the trade-offs that matter, a practical workflow, real numbers, and the questions people ask most — written to be skimmed, applied, and shared.
Key takeaways
- Target the NPU, not just the CPU or GPU, since on modern phones the neural accelerator delivers the best performance-per-watt for sustained inference.
- Keep the model's context and image resolution as low as the task tolerates, because both dominate memory and latency on constrained devices.
- Prefer quantization-aware training or careful post-training quantization with a representative calibration set over naive rounding when accuracy is tight.
- Use the native runtime for the platform you ship on: Core ML on Apple, LiteRT with NNAPI or vendor delegates on Android, and ONNX Runtime for cross-platform.
- For vision-language tasks, pick the smallest VLM that clears your accuracy bar on a benchmark that resembles your real inputs, such as DocVQA for documents.
This is a practical, up-to-date guide to Under the Hood — what it is, why it matters in 2026, and how to apply it in real projects. It is written for developers and founders who want clear answers and proven best practices, not filler.
Whether you're just starting out or leveling up, treat this as a working reference you can return to. Every section is built to be skimmed, applied, and shared.
How vision-language models work
A typical vision-language model (VLM) pairs a vision encoder with a large language model through a projection layer that translates image features into tokens the language model can consume. The vision encoder, historically a CLIP-style or SigLIP transformer, turns an image into a set of patch embeddings, which a small adapter or MLP projects into the LLM's token space. The language model then treats those visual tokens as if they were words, attending over them alongside the text prompt to generate an answer. Architectures such as LLaVA popularized this connector-based recipe, and later designs added higher-resolution tiling and native multimodal pretraining. The elegance is that most of the heavy reasoning still happens in the language backbone, so improvements in LLMs transfer to VLMs.
Mobile AI runtimes: Core ML and LiteRT
Apple's Core ML is the framework for deploying models on iPhone, iPad, and Mac, and it automatically distributes work across the CPU, GPU, and Apple Neural Engine while integrating with tools like coremltools for conversion. On Android, Google's LiteRT, which is the evolution and rebranding of TensorFlow Lite, provides the runtime, with hardware delegates and NNAPI routing operators to vendor NPUs and GPUs. ONNX Runtime offers a cross-platform alternative with execution providers for many accelerators, and Qualcomm, MediaTek, and other silicon vendors ship their own SDKs for their NPUs. Choosing a runtime is mostly about matching the platform you ship on and the accelerators you must reach. Each imposes its own model conversion and operator-support constraints that shape what you can deploy.
TinyML on microcontrollers
TinyML is the practice of running machine learning on microcontrollers with only kilobytes to a few megabytes of RAM and power budgets measured in milliwatts. Typical tasks are always-on and narrow, such as wake-word detection, gesture recognition, predictive maintenance from vibration sensors, and simple anomaly detection. Tooling like LiteRT for Microcontrollers (formerly TensorFlow Lite Micro) and Edge Impulse lets developers train, quantize to 8-bit integers, and deploy models that fit in flash. Because there is no operating system luxury, models are often just a few tens of kilobytes and run without dynamic memory allocation. The appeal is battery-powered or even energy-harvesting devices that can sense and decide locally for months or years.
Small efficient models versus frontier models
Frontier models maximize capability with hundreds of billions of parameters and cloud-scale serving, whereas small efficient models optimize for a fixed footprint of latency, memory, and power. Families such as Gemma, Phi, the smaller Llama variants, Qwen, and Mistral cluster in the 1-to-9-billion-parameter range precisely because that size can run on a phone or laptop while still handling many real tasks. The relevant question is rarely which model is best in the abstract but which is good enough for a specific job within a hard resource budget. Techniques like distillation, pruning, and quantization exist to push more capability into that budget. For narrow, well-scoped tasks, a fine-tuned small model frequently matches a general frontier model at a tiny fraction of the cost.
Edge inference architecture
Edge inference spans a spectrum from powerful phone SoCs down to gateways and microcontrollers, and the right architecture depends on where the device sits on that spectrum. On capable devices the workload is scheduled across CPU, GPU, and a dedicated neural processing unit (NPU), with runtimes dispatching operators to whichever accelerator handles them fastest. Many deployments use a hybrid design where a small local model handles common cases and escalates hard queries to the cloud. Data locality, thermal limits, and battery budget shape these decisions as much as raw accuracy does. Good edge systems also cache aggressively, batch where latency allows, and keep model weights memory-mapped so they load fast and share pages across processes.
Common pitfalls and best practices
The most common mistake is skipping measurement: teams quantize or distill and assume quality held, when only a task-specific evaluation on their own data can confirm it. Another is testing on a desktop and being surprised by thermal throttling, cold-start load times, and missing operator support on the real device. Over-quantizing to 2-bit or 3-bit for the sake of size can quietly wreck reasoning, and feeding VLMs unnecessarily high-resolution images can blow the latency budget for little accuracy gain. Best practice is to build a small held-out benchmark that mirrors production inputs, profile on target hardware early, keep a cloud fallback for hard cases, and treat the quantization level and context length as tunable knobs rather than fixed choices. Version and reproducibility matter too, since a runtime or conversion-tool update can silently change numerics.
Under the Hood: Key Facts and Data
According to recent industry research and the official documentation linked below:
- Industry surveys indicate that privacy, latency, and per-query cost are the three most-cited reasons organizations pursue on-device or edge inference rather than sending every request to a cloud API.
- Knowledge distillation, popularized by Hinton and colleagues in 2015, remains a core technique behind many small production models, with distilled 'student' models often recovering a large share of a much larger teacher's quality.
- TinyML workloads target microcontrollers with kilobytes to low-megabytes of RAM and milliwatt power budgets, enabling always-on tasks such as keyword spotting and anomaly detection on battery- or coin-cell-powered devices.
Quick-Reference Summary
A map of what this guide covers:
| Topic | What you'll learn |
|---|---|
| How vision-language models work | A typical vision-language model (VLM) pairs a vision encoder with a large language model through a projection layer that translates image features into tokens the language model can consume. |
| Mobile AI runtimes: Core ML and LiteRT | Apple's Core ML is the framework for deploying models on iPhone |
| TinyML on microcontrollers | TinyML is the practice of running machine learning on microcontrollers with only kilobytes to a few megabytes of RAM and power budgets measured in milliwatts. |
| Small efficient models versus frontier models | Frontier models maximize capability with hundreds of billions of parameters and cloud-scale serving |
| Edge inference architecture | Edge inference spans a spectrum from powerful phone SoCs down to gateways and microcontrollers |
| Common pitfalls and best practices | The most common mistake is skipping measurement |
How to Get Started with Under the Hood
A simple path that works:
- Learn the fundamentals of Under the Hood from primary sources, not just tutorials.
- Build one small, real project end to end.
- Get feedback, refactor, and add tests.
- Ship it publicly and document what you learned.
- Repeat with a slightly harder project each time.
Build It with a World-Class Full Stack Developer
Sandeep Kumar Chaudhary is a full stack world-class developer. If you want to turn this into a real, production-ready product, get in touch — message directly on WhatsApp at +9779802348957 for a fast, no-pressure consult.
You can also explore the projects already shipped to thousands of users, or start a conversation here.
Final Thoughts
Target the NPU, not just the CPU or GPU, since on modern phones the neural accelerator delivers the best performance-per-watt for sustained inference. The developers and teams who win in 2026 pair strong fundamentals with consistent shipping. Start small, stay curious, build in public, and revisit this guide as your skills grow.
Sources and Further Reading
Frequently Asked Questions
What is under the hood?
Apple's Core ML is the framework for deploying models on iPhone, iPad, and Mac, and it automatically distributes work across the CPU, GPU, and Apple Neural Engine while integrating with tools like coremltools for conversion. On Android, Google's LiteRT, which is the evolution and rebranding of TensorFlow Lite, provides the runtime, with hardware delegates and NNAPI routing operators to vendor NPUs and GPUs. This guide covers under the hood end to end — core concepts, best practices, concrete data, and a step-by-step approach you can apply right away.
Should I use Core ML, LiteRT, or ONNX Runtime?
Use Core ML if you are shipping on Apple devices, since it integrates tightly with the Apple Neural Engine and the iOS and macOS toolchain. Use LiteRT, the successor to TensorFlow Lite, for Android, where delegates and NNAPI reach vendor NPUs. Choose ONNX Runtime when you need one model format that runs across many platforms and accelerators, accepting some per-target tuning.
What is the difference between multimodal AI and a vision-language model?
Multimodal AI is the broad category of models that handle more than one input type, such as text plus images, audio, or video. A vision-language model is a specific and very common kind of multimodal model that combines images and text, typically by pairing a vision encoder with a language-model backbone. Every VLM is multimodal, but multimodal also covers audio, video, and other combinations.
How much accuracy do you lose from quantization?
It depends on the bit width and the method, but 8-bit and well-implemented 4-bit quantization usually preserve most task accuracy, while dropping to 2-bit or 3-bit often degrades reasoning noticeably. Quantization-aware training and careful calibration recover more than naive rounding. The only reliable answer is to benchmark the quantized model on your own task, because losses vary by model and workload.
What is GGUF and why is it everywhere for local LLMs?
GGUF is the file format used by llama.cpp to package quantized language models along with their metadata in a single portable file. It became a de facto standard because llama.cpp runs efficiently on CPUs and consumer GPUs across platforms, and because its graded quant levels let users pick a size-versus-quality point. If you download a local LLM to run on your own machine, it is very likely distributed as a GGUF file.
Sandeep Kumar Chaudhary
Full Stack Software Developer· Nepal's SEO, AEO, GEO & AIO expert and share-market educator. More about me