AUDIO / SPEECH TRANSCRIPTION

Tock

Good words. Worth keeping.

Turn English speech into text with compact, locally running speech recognition.

Try Tock Download weights ↓

LanguageEnglish

RunsLocally

ModelWhisper tiny.en

ReleaseResearch v0.2

A LITTLE LESS SETUP

A recording in.
A transcript out.

Press play. Then give Tock a listen.
No account. No audio upload.

TOCK / LIVE DEMO Audio stays on your device

01 / YOUR RECORDING

A few words that made history.

JFK’s inaugural address · short sample

Choose an audio file

English · up to 30 seconds · 15 MB max
WAV, MP3, or another format your browser can read.

First run downloads 32.2 MB of weights plus the browser runtime. Saved locally when browser storage is available.

02 / THE WORDS

A little audio.
Something you can read.

Your transcript will appear here.

Ready when you are.

Same Tock Q5_1 weights. Browser CPU preview; speed depends on your device. Review important transcripts.

Sample: John F. Kennedy, 1961 inaugural address, public domain · Listen to the sample

SPEECH, CLOSE TO HOME

From voice
to something useful.

Voice notes, recorded ideas, and spoken drafts. Tock explores compact English transcription that runs on your own machine.

The current candidate uses OpenAI’s pretrained Whisper tiny.en, converted by the whisper.cpp project and quantized locally to Q5_1. Our work here is model compression, runtime integration, and comparative evaluation.

Research candidate · Try the browser CPU preview above, or use the native local runtime. Android ASR integration is not released yet.

ORIGINAL VS COMPRESSED

58.6% smaller.

Same recordings. Same decoder.
Same text normalization.

32.2 MBOn-device weights

58.6% smaller than the 77.7 MB original GGML artifact.

6.20%Word error rate

Full LibriSpeech test-clean. Original FP16: 5.57%.

Whisper tiny.en · LibriSpeech test-clean · lower WER is better
Measurement	Original FP16	Tock Q5_1
Word error rate	5.57%	6.20%
Weights on disk	77.7 MB	32.2 MB
Evaluation wall time	273.0 s	265.8 s

All 2,620 recordings, 5.40 hours of read English speech. M4 Pro CPU, four threads, Apple Accelerate, greedy decoding, no temperature fallback, no prior text context. Whisper’s English normalizer is applied to both outputs and references. Wall time includes batched process startup, loading, and output writing; one pass per variant, not a repeated latency benchmark or phone measurement. These are our runs, not imported leaderboard scores.

UNDER THE SURFACE

Specs.

Speed: 5.40 hours of test audio in 265.8 seconds on an M4 Pro, four CPU threads. Single batch evaluation including model loading; not a streaming or device latency benchmark.
Accuracy: 6.20% word error rate on the complete LibriSpeech test-clean split. See the original-model comparison above.
Model: OpenAI Whisper tiny.en, 39M parameters. Encoder–decoder Transformer; GGML conversion from whisper.cpp; local Q5_1 quantization.
On-device size: 32.2 MB GGML weights. Runtime and working memory are additional.
Languages: English transcription. Translation and multilingual recognition are outside this release.
Input: Native runtime: 16 kHz mono PCM WAV. Browser preview: audio files up to 30 seconds, decoded and resampled locally.
Output: Transcript text in the browser preview. Native CLI also provides segment timestamps. Word-level timestamp accuracy is not benchmarked.
Platforms: Native macOS CPU runtime with Node.js and Kotlin/JVM adapters. Browser CPU preview using WebAssembly; desktop recommended. Android ASR integration is not released.
License: MIT upstream weights and runtime, with original attribution retained.

ON YOUR MACHINE

Take Tock
into your own project.

Node.js and Kotlin/JVM adapters require a local runtime and model weights. Packages are not yet published to registries.

import { transcribe } from '@embermote/tock-asr';

const result = await transcribe({
  executablePath: 'whisper.cpp/build/bin/whisper-cli',
  modelPath: 'tock-q5_1.bin',
  audioPath: 'recording.wav',
});
console.log(result.text);
// Native Node.js. Local runtime and weights required.

Local artifact → local inferenceNo API key.

MODEL DETAILS

Know what you’re running.

Pretraining comes from upstream Whisper. This candidate has been quantized, not newly pretrained or fine-tuned. Test-clean is evaluation data only; no recordings are published here.

Model card Benchmark report Model license Runtime license Upstream model card Native runtime Evaluation dataset

Read-speech results do not establish accuracy on meetings, accents, noisy microphones, or long-form recordings. Review important transcripts.