speech-analyzer-cli

A small macOS command-line interface for Apple's on-device SpeechAnalyzer and SpeechTranscriber APIs. It transcribes prerecorded audio and can emit plain text, word-level JSON, JSONL, SubRip, or WebVTT.

Built using GPT-5.6 Sol high in ChatGPT followed by GPT-5.6 Sol xhigh in Codex CLI to fix some bugs.

• macOS 26 or later
• Apple silicon supported by SpeechTranscriber
• Xcode 26 or the matching Command Line Tools
• A network connection the first time each language model is installed

chmod +x scripts/install.sh
./scripts/install.sh

This builds a background application bundle beneath ~/.local/libexec and links its executable as ~/.local/bin/speech-analyzer. The bundle supplies the privacy metadata macOS needs when requesting Speech Recognition permission.

If necessary, add this to your shell configuration:

export PATH="$HOME/.local/bin:$PATH"

The first transcription should prompt for Speech Recognition permission. The audio itself is transcribed on-device. A language model may be downloaded and installed into system-managed storage before the first transcription.

speech-analyzer interview.m4a
speech-analyzer --locale en-GB --format json interview.m4a
speech-analyzer --format jsonl interview.wav > words.jsonl
speech-analyzer --format srt --output interview.srt interview.m4a
speech-analyzer --format vtt --output interview.vtt interview.m4a
speech-analyzer --list-locales

JSON contains the input path, selected locale, complete text, and an array of word objects:

{
  "file": "/path/to/interview.m4a",
  "locale": "en-US",
  "text": "Hello from SpeechAnalyzer.",
  "words": [
    {"start": 0.42, "end": 0.81, "text": "Hello", "confidence": 0.94}
  ]
}

jsonl writes one word object per line. Confidence is omitted by Swift's Codable encoder when Apple does not supply it.

SRT and WebVTT cues are grouped using sentence punctuation plus conservative word-count, character-count, and duration limits. SpeechTranscriber's word ranges are useful for playback synchronization, but should not be interpreted as precise forced-alignment boundaries around silence.

swift test
swift run speech-analyzer --help

The formatting code lives in the platform-independent SpeechAnalyzerCore target. Actual transcription can only be built and exercised on macOS 26 with Apple's Speech framework.

If permission was denied, enable the tool under System Settings → Privacy & Security → Speech Recognition.

The installer uses an ad-hoc signature. Rebuilding may cause macOS to ask for permission again. For a stable development installation, sign the app bundle with an Apple Development certificate instead.

If a locale is rejected, compare it with:

speech-analyzer --list-locales

Locale matching currently uses exact, case-insensitive BCP-47 identifiers.