@modelcontextprotocol/server-transcript

Exposes real-time speech-to-text transcription as an MCP server resource, allowing Claude and other MCP clients to subscribe to and consume live audio transcription streams. Implements the MCP protocol's resource subscription model to push transcribed text segments as they become available, with support for streaming audio input from system audio devices or network sources.

Implements MCP's resource streaming interface to deliver transcribed audio segments incrementally to clients as they complete. Uses the MCP protocol's resource URI scheme and subscription mechanism to manage client connections, handle backpressure, and ensure reliable delivery of transcript chunks without requiring clients to poll or manage connection state.

Captures audio from system audio devices (microphone, line-in, or virtual audio devices) and forwards it to the transcription engine. Handles audio format negotiation, sample rate conversion, and device enumeration to allow users to select input sources. Likely uses Node.js audio libraries (e.g., node-portaudio, naudiodon) to interface with OS-level audio APIs.

Normalizes incoming audio streams to a standard format (likely 16-bit PCM at 16kHz) required by the transcription engine. Handles sample rate conversion, bit depth adjustment, and channel mixing (stereo to mono) transparently. Uses audio resampling algorithms to maintain quality during format conversion without requiring client-side preprocessing.

Abstracts the underlying speech-to-text engine behind a provider interface, allowing selection of different transcription backends (e.g., Web Speech API, Whisper, Google Cloud Speech-to-Text, Azure Speech Services). Likely implements a plugin or strategy pattern to swap transcription providers without changing server code. Handles API authentication, error handling, and fallback logic.

Buffers transcribed text segments and manages delivery timing to MCP clients, balancing latency (pushing segments as soon as available) with throughput (batching small segments to reduce overhead). Implements configurable buffering strategies (e.g., time-based, size-based, or confidence-based) to control when transcript chunks are sent to clients. Handles partial transcripts (interim results) vs. final transcripts.

Manages MCP server initialization, shutdown, and resource cleanup. Implements MCP server protocol handshake, handles client connections and disconnections, and ensures graceful shutdown of audio capture and transcription pipelines. Likely uses MCP SDK for Node.js to handle protocol details and resource registration.