Your comprehensive reference for AI, speech recognition, LLM, and developer terminology. Learn the terms that power modern voice-first development.
70 terms defined
An AI agent is an autonomous software entity that can perceive its environment, make decisions, and take actions to achieve specific goals. In the context of AI coding tools, agents can execute multi-step tasks, use tools, and interact with codebases.
AI agents enable developers to delegate complex, multi-step coding tasks. Unlike simple chatbots, agents can plan, execute, and iterate on solutions autonomously.
A unique identifier used to authenticate requests to an API (Application Programming Interface). API keys are used to track usage, manage access, and bill for services.
Securely managing API keys is essential for using cloud AI services. Leaking keys can lead to unauthorized charges and data breaches.
Technology that converts spoken language into written text. ASR systems analyze audio input, identify phonemes and words, and produce text transcriptions.
ASR is the foundation of voice-first applications. High-quality ASR enables natural voice interactions without requiring users to type.
A neural network component that allows models to focus on relevant parts of input data. The transformer architecture uses self-attention to weigh the importance of different input tokens.
Attention mechanisms enabled the creation of large language models. They allow models to understand context and relationships across long sequences of text.
AI systems capable of operating independently to achieve goals without constant human intervention. These systems can make decisions, take actions, and adapt to changing conditions.
Autonomous AI agents can handle complex workflows, freeing developers to focus on high-level decisions rather than repetitive tasks.
Processing multiple items as a group rather than individually. In AI, batch processing allows efficient handling of multiple prompts or transcriptions simultaneously.
Batch processing reduces latency and costs when working with large volumes of data. It is essential for efficient audio processing pipelines.
A model where users provide their own API keys for cloud services rather than using shared keys. This gives users control over their usage and billing.
BYOK provides transparency and control. You pay only for what you use and can choose your preferred AI provider.
A prompting technique where the AI is encouraged to show its reasoning step-by-step before reaching a conclusion. This often improves accuracy on complex tasks.
Chain of thought prompting dramatically improves AI performance on reasoning tasks. It helps models break down complex problems into manageable steps.
An AI assistant created by Anthropic, known for being helpful, harmless, and honest. Claude is designed with strong safety practices and is popular among developers.
Claude excels at coding tasks and nuanced reasoning. Its large context window and safety-focused design make it ideal for professional development work.
AI services hosted on remote servers, accessed via the internet. Cloud AI offers powerful models without requiring local hardware but requires an internet connection.
Cloud AI provides access to cutting-edge models that would be impossible to run locally. However, it raises privacy considerations for sensitive data.
The maximum amount of text (measured in tokens) that an LLM can process at once. Larger context windows allow models to consider more information when generating responses.
Context window size determines how much code or conversation history an AI can consider. Larger windows enable more comprehensive understanding of complex codebases.
An AI-powered code editor built on VS Code that integrates LLMs directly into the development workflow. It enables natural language coding and intelligent code completion.
Cursor represents the future of AI-assisted development, where natural language becomes a first-class way to write and edit code.
The process of partitioning an audio stream into segments according to speaker identity. Diarization determines "who spoke when" in a recording.
Speaker diarization is essential for transcribing meetings and conversations accurately. It enables attribution of speech to specific individuals.
A technique for training smaller models to replicate the behavior of larger models. The smaller "student" model learns from the larger "teacher" model.
Distillation enables running powerful AI capabilities on resource-constrained devices. It makes local AI practical for everyday use.
A numerical representation of data (text, audio, images) as a vector in high-dimensional space. Similar items have embeddings that are close together.
Embeddings enable semantic search and similarity matching. They power RAG systems that retrieve relevant context for AI responses.
A system architecture where a single model handles the entire pipeline from input to output. In speech recognition, end-to-end models process audio directly to text.
End-to-end systems like Whisper are simpler and often more accurate than traditional multi-stage pipelines.
A prompting technique where examples are provided to guide the AI's response format and behavior. This helps models understand the desired output pattern.
Few-shot learning makes AI more reliable and predictable. It is essential for getting consistent, well-formatted outputs.
The process of further training a pre-trained model on specific data to improve performance on particular tasks. Fine-tuning adapts general models to specialized use cases.
Fine-tuning can dramatically improve AI performance for specific domains. It enables customization without training from scratch.
A file format for storing quantized LLM models, designed for efficient loading and inference. GGUF replaced the older GGML format and is used by llama.cpp.
GGUF enables running large models on consumer hardware. It is the standard format for local AI deployment.
A family of large language models developed by OpenAI. GPT models are trained on vast text data and can generate human-like text, answer questions, and assist with coding.
GPT models revolutionized AI capabilities and spawned the current generation of AI assistants and coding tools.
A company that designs custom AI chips (LPUs) optimized for inference. Groq's cloud service offers extremely fast LLM inference speeds.
Groq's speed makes real-time AI interactions feel instantaneous. Low latency is crucial for voice-first applications.
When an AI model generates information that is factually incorrect, inconsistent, or entirely fabricated. Hallucinations occur when models confidently produce plausible but false content.
Understanding hallucinations is critical for using AI safely. Always verify AI-generated code and claims against authoritative sources.
Technology that listens for specific trigger words or phrases to activate a voice interface. Examples include "Hey Siri" or "OK Google."
Hotword detection enables always-on voice interfaces without constantly processing speech. It balances responsiveness with privacy and efficiency.
A software application that provides comprehensive tools for software development, including code editing, debugging, and build automation.
Modern IDEs with AI integration transform the development experience. Tools like Cursor and VS Code with AI extensions accelerate coding significantly.
The process of running a trained AI model to generate predictions or outputs. Inference is what happens when you send a prompt to an LLM and receive a response.
Inference speed and cost directly impact user experience. Fast inference enables real-time AI interactions.
The ability of LLMs to learn from examples provided within the prompt, without updating model weights. This enables task adaptation through prompting alone.
In-context learning makes LLMs incredibly flexible. You can teach new behaviors through examples rather than retraining.
A configuration for LLMs that ensures outputs are valid JSON format. This is essential for programmatic consumption of AI responses.
JSON mode eliminates parsing errors when integrating AI into applications. It enables reliable structured data extraction.
The date after which an AI model has no knowledge of events or information. Models are trained on data up to their cutoff date.
Understanding knowledge cutoff helps set expectations. For current information, use RAG or web-connected AI tools.
The delay between sending a request and receiving a response. In AI applications, latency includes network time, model processing, and response generation.
Low latency is essential for voice-first applications. High latency breaks the flow of natural voice interaction.
A type of artificial intelligence model trained on vast amounts of text data. LLMs can understand and generate human-like text, enabling applications like chatbots, code completion, and content generation.
LLMs are the foundation of modern AI coding tools. Understanding their capabilities and limitations is essential for effective use.
An architectural approach where applications process data locally on the user's device rather than sending it to remote servers. Local-first prioritizes privacy and offline capability.
Local-first AI keeps your code and voice data private. It works offline and eliminates cloud dependency.
A large language model that runs entirely on your local device without requiring internet connectivity. Local LLMs provide privacy and work offline.
Local LLMs keep your data private and work without internet. They are essential for sensitive codebases and air-gapped environments.
A technique for efficiently fine-tuning large models by training only a small number of additional parameters. LoRA adapters can be swapped to change model behavior.
LoRA makes fine-tuning accessible without massive compute resources. It enables personalized AI without full model retraining.
A protocol developed by Anthropic for connecting AI models to external data sources and tools. MCP standardizes how AI assistants access context.
MCP enables AI assistants to work with your specific tools and data. It is becoming a standard for AI integration.
A visual representation of audio frequency content over time, using the mel scale that approximates human perception. Mel spectrograms are the input format for many speech models.
Understanding mel spectrograms helps when debugging audio processing pipelines. They are the bridge between raw audio and AI models.
The number of parameters in a neural network, typically measured in billions (e.g., 7B, 70B). Larger models generally have more capability but require more resources.
Model size affects quality, speed, and hardware requirements. Choose the right size for your balance of capability and resources.
A computing system inspired by biological brains, consisting of interconnected nodes (neurons) that process information. Neural networks are the foundation of modern AI.
Neural networks enable machines to learn patterns from data. They power everything from speech recognition to code generation.
A field of AI focused on enabling computers to understand, interpret, and generate human language. NLP powers applications from translation to chatbots.
NLP makes AI interaction natural and accessible. You can communicate with AI using everyday language rather than code.
An open-source tool for running large language models locally on your computer. Ollama simplifies the process of downloading, managing, and running local AI models.
Ollama makes local AI accessible to everyone. It eliminates the complexity of setting up local LLM infrastructure.
An AI research company that created GPT, ChatGPT, Whisper, and other influential AI technologies. OpenAI's models are widely used in commercial applications.
OpenAI's models set industry benchmarks. Understanding their offerings helps choose the right tools for your needs.
A unified API gateway that provides access to multiple AI model providers through a single interface. OpenRouter simplifies model comparison and switching.
OpenRouter enables using the best model for each task without managing multiple API keys. It provides flexibility and fallback options.
The learned values within a neural network that determine its behavior. During training, parameters are adjusted to improve model performance.
Parameter count is a rough proxy for model capability. More parameters generally mean more capacity to learn complex patterns.
The initial phase of training an AI model on large amounts of general data. Pre-training creates a foundation that can be fine-tuned for specific tasks.
Pre-training is what makes LLMs so capable. The vast knowledge acquired during pre-training enables diverse applications.
In Vocoding, a profile is a configuration that defines how prompts should be optimized for a specific use case or target AI system.
Profiles enable consistent, optimized prompts without manual adjustment. Different profiles work best for different AI tools.
The input text given to an AI model to guide its response. Prompts can include instructions, context, examples, and constraints.
The quality of your prompt directly affects AI output quality. Good prompts are clear, specific, and well-structured.
The practice of designing and optimizing prompts to get the best results from AI models. Prompt engineering involves understanding model behavior and crafting effective instructions.
Prompt engineering can dramatically improve AI output quality. It is a core skill for effective AI use.
The process of automatically improving prompts for better AI results. This includes restructuring, adding context, and formatting for specific AI systems.
Prompt optimization automates best practices. It saves time and improves consistency compared to manual prompt crafting.
A technique that reduces model size by using lower precision numbers (e.g., 4-bit instead of 16-bit). Quantization enables running larger models on limited hardware.
Quantization makes local AI practical on consumer hardware. It trades minimal accuracy for major memory savings.
A technique that combines information retrieval with LLM generation. RAG systems fetch relevant documents and include them in the prompt context.
RAG grounds AI responses in your specific data. It reduces hallucinations and enables AI to work with current information.
Restrictions on how many API requests can be made within a time period. Rate limits prevent abuse and ensure fair resource distribution.
Understanding rate limits helps design reliable applications. Hitting limits can disrupt workflows and require fallback strategies.
Converting speech to text as it is spoken, with minimal delay. Real-time transcription enables live captioning and voice interfaces.
Real-time transcription enables natural voice workflows. Low latency is essential for maintaining conversational flow.
A mechanism where a model relates different positions within the same sequence to compute a representation. Self-attention is the core of transformer architecture.
Self-attention enables models to understand relationships across entire documents. It is what makes modern LLMs so capable.
Search based on meaning rather than exact keyword matching. Semantic search uses embeddings to find conceptually related content.
Semantic search finds relevant information even when exact terms differ. It enables more intelligent code and document retrieval.
Technology that converts spoken language into written text. Also known as speech recognition or automatic speech recognition (ASR).
Speech-to-text enables voice-first workflows. High-quality STT is the foundation of natural voice interfaces.
Delivering content progressively as it is generated rather than waiting for completion. Streaming shows partial results immediately.
Streaming improves perceived latency. Seeing results appear progressively feels faster than waiting for a complete response.
The process and technology of converting spoken audio into written text. STT is a fundamental component of voice-enabled applications.
STT is the bridge between voice and text-based AI. Accurate STT is essential for voice-first development workflows.
Instructions given to an AI model that define its role, behavior, and constraints. System prompts set the context for all subsequent interactions.
System prompts shape AI behavior consistently. They are essential for creating reliable, predictable AI applications.
A parameter that controls the randomness of AI model outputs. Lower temperatures produce more focused, deterministic responses; higher temperatures increase creativity and variation.
Temperature tuning balances consistency and creativity. Lower temperatures are better for code; higher for brainstorming.
The basic units of text that LLMs process. Tokens can be words, subwords, or characters depending on the tokenizer. One token is roughly 4 characters in English.
Understanding tokens helps estimate costs and context limits. Token efficiency affects both speed and price of AI operations.
The ability of AI models to invoke external functions or APIs. Tool calling enables AI to take actions beyond text generation, such as running code or searching the web.
Tool calling transforms AI from a text generator into an active assistant. It enables AI to interact with systems and execute tasks.
The conversion of audio content into written text. Transcription can be real-time or batch, and may include timestamps and speaker identification.
Accurate transcription is the first step in voice-to-code workflows. Quality transcription directly affects downstream AI performance.
A neural network architecture introduced in 2017 that uses self-attention mechanisms. Transformers are the foundation of modern LLMs and speech models.
The transformer architecture enabled the current AI revolution. Understanding it helps comprehend AI capabilities and limitations.
Technology that converts written text into spoken audio. Modern TTS systems can produce natural-sounding speech with appropriate intonation and emotion.
TTS enables audio feedback and accessibility features. Combined with STT, it creates complete voice interaction loops.
Technology that detects when speech is present in an audio stream. VAD distinguishes speech from silence, background noise, and other non-speech sounds.
VAD enables efficient audio processing by only transcribing speech segments. It reduces costs and improves accuracy.
A database optimized for storing and querying high-dimensional vectors (embeddings). Vector databases enable efficient similarity search at scale.
Vector databases power RAG systems and semantic search. They are essential for AI applications that work with large document collections.
A design philosophy that prioritizes voice interaction as the primary input method. Voice-first applications are designed around speaking rather than typing.
Voice-first workflows can be faster and more natural than typing. They enable hands-free operation and accessibility.
Video RAM - the memory on a graphics card used for storing data needed for rendering and computation. Running local LLMs requires sufficient VRAM.
VRAM limits which models you can run locally. Understanding VRAM requirements helps choose appropriate hardware and model sizes.
The numerical values that define a neural network's learned behavior. Weights are adjusted during training and determine model outputs.
Model weights represent learned knowledge. Downloading weights is how you get a pre-trained model onto your system.
An open-source speech recognition model developed by OpenAI. Whisper provides high-accuracy transcription in multiple languages and can run entirely on local hardware.
Whisper enables private, local speech recognition without sending audio to the cloud. It is the foundation of privacy-first voice applications.
The ability of AI models to perform tasks without any task-specific examples. Zero-shot relies entirely on the model's pre-trained knowledge and clear instructions.
Zero-shot capability means AI can help with novel tasks immediately. It reduces the need for custom training or extensive examples.
Vocoding combines these technologies to help you code faster with voice-first AI.