Free Audio to Text — Whisper Transcription in Your Browser

Transcription without the upload

Every mainstream cloud transcription service follows an identical architecture: you upload your audio recording to a remote server cluster, it gets transcribed somewhere in the cloud, and the text is returned over the network. While that workflow may be acceptable for a public podcast episode, it presents serious confidentiality risks for a doctor-patient consultation, an executive board meeting, an unreleased investigative interview, or a voice memo containing personal financial details. The spoken voice is among the most sensitive personal data types, yet standard transcription pipelines route it through third-party servers by default.

YuliusBox inverts this model. OpenAI's Whisper model executes directly inside your browser tab through WebAssembly and the ONNX runtime via @xenova/transformers. The audio data never crosses the network; transcription runs entirely on your device's local CPU, and the resulting transcript appears on your screen in real time. For journalists protecting confidential sources, legal practitioners reviewing case depositions, academic researchers bound by strict ethics agreements, and privacy-conscious professionals, this on-device approach provides verifiable data sovereignty.

Verified client-side pipeline: 16 kHz PCM decoding, VAD silence trimming, and 30s/5s chunking

Instead of relying on unverifiable single-sentence speed or accuracy claims, you can inspect the exact deterministic signal-processing pipeline executed by AudioToTextTool.tsx and public/whisper.worker.js on every recording:

1. 16,000 Hz mono PCM normalization: When you select an MP3, WAV, or M4A file (or stop a MediaRecorder microphone capture), the browser's Web Audio API (new AudioContext({ sampleRate: 16000 }).decodeAudioData) resamples the audio into a single-channel 32-bit floating-point array (Float32Array). Because each sample occupies 4 bytes at 16,000 samples per second, decoded audio consumes an exact, predictable 64,000 bytes (62.5 KiB) per second — or 3,840,000 bytes (3.66 MiB) per minute — in local tab memory.

2. Energy-based Voice Activity Detection (trimSilence): Before invoking the neural network, our pre-processor scans the 16 kHz Float32Array in 20 ms windows (320 samples) and identifies active speech where mean absolute amplitude exceeds 0.02. It trims leading and trailing silence while retaining a 250 ms guard margin (4,000 samples) on both ends so initial consonants are never clipped, and bypasses trimming entirely if effective speech is shorter than 0.5 seconds (8,000 samples).

3. Sliding-window ONNX inference: Inside the Web Worker, @xenova/transformers v2 executes the quantized 8-bit (q8) Whisper encoder and merged decoder graphs with task: 'transcribe', chunk_length_s: 30, and stride_length_s: 5. Recordings longer than 30 seconds are segmented into overlapping 30-second windows with a 5-second stride on each boundary so words spoken across chunk transitions are reconciled cleanly.

Realistic accuracy expectations: While a clean, close-mic single-sentence English voice memo routinely transcribes without word errors on the default Base model, single-sentence read-speech tests do not predict Word Error Rate (WER) on noisy multi-speaker meetings, overlapping crosstalk, or reverberant conference rooms. For accented speech, domain jargon, or room echo, switch from Base to the Small tier and explicitly select the spoken language rather than leaving the dropdown on Auto Detect.

How YuliusBox compares to other browser-based Whisper transcribers

Running OpenAI Whisper in the browser is a shared architecture across several privacy-first transcription tools, including Whisper Web (whisperwebfree.com), Ryusei Transcriber (transcriber.tools.ryusei.io), Pageonaut, and Floi. On core privacy — keeping audio files on-device with zero server uploads — all of these tools tie.

Where competitors offer more features: Whisper Web and Ryusei Transcriber support WebGPU hardware acceleration alongside WebAssembly and export timestamped subtitle formats (.srt, .vtt, .json), and Ryusei Transcriber includes on-device speaker diarization and multi-gigabyte models (up to 5.2 GB). YuliusBox currently executes CPU WebAssembly via ONNX Runtime Web and exports plain text (.txt) and clipboard copy.

Where YuliusBox differs: First, we pre-process decoded 16 kHz audio through an automatic energy-based Voice Activity Detection (trimSilence) pass that strips leading and trailing dead air (while keeping a 250 ms guard buffer) and flags recordings under 3.0 seconds before Whisper's transformer decoder can hallucinate repeated phrases on silence. Second, all three quantized q8 ONNX model tiers (43.34 MB Tiny, 79.39 MB Base, and 251.59 MB Small) are served from our versioned asset CDN with automatic tier fallback to Base if a download is interrupted. Third, you can verify zero-upload execution in DevTools (F12 -> Network): after the initial GET requests for the .onnx model weights and ort-wasm runtime complete, toggling the browser to Offline mode and uploading an MP3 produces zero outbound network packets while transcription runs in the local Web Worker.

Technical limitations, model sizes, and hardware requirements

Because Whisper models execute in client-side Web Workers using WebAssembly rather than dedicated cloud GPUs, users should be aware of several practical constraints:

Model download sizes and memory usage: The tool offers three quantized 8-bit (q8) ONNX model tiers. Tiny requires an initial download of 43.34 MB (41.33 MiB) and ~150 MB of browser RAM, making it exceptionally fast for rough drafts. Base downloads 79.39 MB (75.71 MiB) and uses ~250 MB of RAM, offering the optimal balance of accuracy and speed for most audio. Small downloads 251.59 MB (239.93 MiB) and requires ~600 MB of RAM, delivering high accuracy on technical jargon and thick accents. Model weights download once and are cached persistently in browser storage (CacheStorage and IndexedDB), allowing subsequent transcriptions to run even completely offline.

Hardware recommendations: The Tiny and Base tiers run smoothly on modern laptops, desktops, and recent smartphones. The Small tier is computationally intensive; mobile phones or computers with less than 4 GB of available system memory may experience high CPU usage or browser tab throttling.

Short clip advisory: Audio clips under 3.0 seconds trigger an in-app guidance warning. Whisper's transformer encoder relies on acoustic and semantic context; very short isolated sounds or single words can occasionally cause the model to repeat phantom phrases or skip words. Speaking a full, continuous sentence dramatically improves recognition accuracy.

Supported audio formats and export limits: The tool decodes MP3, WAV, and M4A/AAC files natively, as well as live microphone streams via the browser's MediaRecorder API. Export is currently limited to plain text (.txt) and direct clipboard copying; timestamped subtitle formats (.srt or .vtt) are planned for a subsequent update.

Choosing a model tier: tiny vs base vs small

The three model sizes represent a deliberate trade-off between download footprint, execution speed, and linguistic accuracy. Tiny is the fastest option: it loads almost instantaneously on standard mobile connections and transcribes clear, single-speaker English with minimal latency. It is ideal for quick voice dictation, meeting minutes, and brief personal reminders.

Base is the default configuration recommended for 90% of use cases. It demonstrates substantially stronger resilience against background ambient noise, conversational crosstalk, and non-native accents, while transcribing at roughly 3x to 5x real-time speed on modern consumer hardware. Small is the precision instrument: it delivers the lowest word error rate on difficult recordings featuring overlapping dialogue, medical or technical terminology, or reverberant conference rooms, but requires a capable multi-core computer.

Getting accurate transcripts: recording matters more than settings

In speech-to-text systems, acoustic signal quality is far more decisive than model parameter count. A clean recording processed by the Tiny model will consistently outperform a muffled, echoey recording processed by the Small model. Always position the microphone within 15–30 centimeters of the speaker's mouth, and avoid placing a phone flat on the center of a large table in speakerphone mode.

For interviews or remote calls, using basic wired earphones with an inline microphone dramatically reduces room reverberation compared to built-in laptop microphones. If you are uploading existing recordings, avoid re-compressing them to low-bitrate MP3s beforehand; Whisper processes raw audio waveforms natively. Finally, explicitly selecting your target language in the dropdown rather than relying on Auto Detect improves accuracy on brief recordings.

What you can do with on-device transcripts

Local transcription unlocks productivity workflows that were previously impossible due to privacy policies or SaaS subscription costs. Professionals can transcribe confidential executive standups, legal interview tapes, and medical consultation notes without violating non-disclosure agreements or GDPR regulations.

Content creators and podcasters can turn audio recordings into searchable text archives, draft blog summaries, or produce initial video script outlines. Because there are no monthly minute limits, no paywalls, and no subscription tiers, transcribing an hour-long lecture costs exactly the same as transcribing a ten-second voice note: completely free.