ブラウザ内で完結するローカルAI音声文字起こし（Whisper WebAssembly）

Why In-Browser Speech Recognition Changes Audio Editing

Spoken-word post-production traditionally required two disconnected tools: a cloud transcription API that charged per minute and uploaded private voice recordings to third-party servers, and a separate digital audio workstation (DAW) where editors manually cut pauses and aligned subtitles by hand.

Browser-local speech recognition eliminates both bottlenecks. By running an optimized transformer speech model directly in client WebAssembly, creators transcribe podcasts, video voiceovers, and interviews in seconds with zero subscription costs, zero queue latency, and 100% data privacy.

How Whisper WebAssembly Executes in Browser Memory

When you click Transcribe with AI, the browser initializes a background Web Worker that loads quantized ONNX weights of OpenAI’s Whisper model. The Web Audio engine converts your recording into 16,000 Hz 32-bit floating-point mono audio and streams chunks into the transformer encoder.

Because computation runs on your device CPU/GPU via WebAssembly SIMD instructions, confidential interviews, corporate memos, and unreleased podcast episodes never leave your local machine.

Word-Level Timestamp Tokenization vs Sentence Blocks

Standard video players only require sentence-level cues. However, surgical audio editing requires word-level token timestamps. Every word token contains:

This sub-second timestamp granularity enables automated algorithms to pinpoint the exact start and end of verbal pauses, hesitation crutches (um, uh), and repeated stumbles.

Step-by-Step Workflow: Speech-to-Text to Synchronized Master

ステップ 1：Decode Audio Locally via Web Audio API

Drop your MP3, WAV, M4A, or MP4 file into the workspace. The browser’s native AudioContext.decodeAudioData() decodes the container into uncompressed multi-channel PCM buffers in milliseconds.

ステップ 2：Execute Local Whisper Transcription in a Web Worker

Click Transcribe with AI in the Subtitles tab. The background worker tokenizes acoustic features, emits progressive percentage updates, and returns structured subtitle cues.

ステップ 3：Scan Transcript for Fillers, Stutters, and Retakes

The NLP engine analyzes the transcript against conversational hesitation lexicons, flags repeated-word stutters, and clusters duplicate spoken phrases into winning takes and stumble outtakes.

ステップ 4：Apply Sentence Protection to Retain Natural Pauses

Unlike destructive noise gates that cut dialogue gaps blindly, sentence protection recognizes active speech cues and retains a 0.32s speech cushion, keeping conversational rhythm natural and human.

ステップ 5：Export Synchronized Subtitles and Lossless WAV

Export your finished audio as a 16-bit PCM WAV master alongside perfectly time-aligned WebVTT (.vtt) and SubRip (.srt) subtitle tracks where cue timecodes have been mathematically shifted across every edited cut.

Local WebAssembly Whisper vs Cloud Transcription APIs

音声編集・処理で避けるべき代表的な失敗と注意点

1. 1. Editing audio after generating subtitles without timecode recalculation: Always use an editor that recalculates cue offsets automatically when pauses are trimmed.
2. 2. Deleting all hesitation sounds indiscriminately: Subtle pauses convey empathy and nuance; preserve 0.32s breathing cushions rather than hard-cutting to zero.
3. 3. Exporting lossy MP3s between editing steps: Always export uncompressed 16-bit PCM WAV to prevent generational compression artifacts.
4. 4. Ignoring repeated take stumbles: Check the Retake drawer to audition flubbed sentences against your final winning take before final export.
5. 5. Uploading sensitive corporate memos to cloud speech APIs: Use local WebAssembly speech recognition to guarantee NDA compliance and client confidentiality.