Running Local AI Speech-to-Text in Browser CPU with WebAssembly: Zero Server Uploads

Uploading sensitive audio files to remote cloud speech-to-text APIs poses a significant privacy concern. Leading AI transcription companies require users to send their voice data to their servers, leaving it vulnerable to breaches or model retraining. However, with the advent of WebAssembly SIMD and ONNX Runtime Web, it's now possible to execute quantized Whisper transformer models entirely within a user's browser, thanks to SolveMyMedia Transcribe.

SolveMyMedia Transcribe leverages Transformers.js to decode audio into 16kHz float buffers in browser memory and processes them directly through a local model runtime. To get started, import the pipeline function from @xenova/transformers and initialize the Whisper Tiny English model using the Xenova/whisper-tiny.en endpoint, specifying the webgpu device for computation.

Then, use the transcribeLocalAudio function to pass an audioBlob, which is converted to an arrayBuffer and passed to the transcriber. The function returns the transcription text, ensuring that no data leaves the user's machine.

When comparing the privacy and cost implications, the Cloud Speech API, provided by major services like OpenAI or Rev, stores audio files on external cloud servers, incurring a cost of $0.006 per minute. In contrast, SolveMyMedia Local AI Transcribe keeps all data within the client's RAM, incurring zero costs and ensuring complete privacy. Additionally, the local solution supports offline operation without internet dependency, unlike the cloud-based API, which requires an internet connection.

Try out SolveMyMedia Transcribe yourself at https://solvemymedia.com/transcribe and share your experiences with client-side AI inference in production in the comments section!