Fixes SIGTRAP on ohr --listen (#1) via SpeechAnalyzer.bestAvailableAudioFormat + AVAudioConverter. TCC keys in Info.plist. Locale fallback. Thanks @pallas.

brew upgrade ohr

Contributors

• @pallas

Comprehensive QA Report

What was tested

• Formats: m4a, wav, mp3, aiff, flac — all 5 work
• Durations: 5s, 30s, 1min, 3min, 10min — all work, no upper limit found
• Performance: ~50x real-time on Apple M2 (10 min audio → 4.7 seconds)
• Edge cases: empty files, corrupted data, silence, wrong format — all handled cleanly
• Languages: 30 reported, English verified, others need real speech testing
• Server: All endpoints, all response formats, security, CORS — all passing

How it was tested

All tests used synthetic speech from macOS say command. This is an important caveat documented honestly in the README and docs/testing.md. Real human speech with accents, noise, and natural pauses has NOT been tested and will likely produce lower accuracy.

Known limitations documented

• ~90-95% accuracy on clear synthetic speech
• Number/ordinal confusion ("five second" → "52nd")
• No speaker diarization
• apfel context window limits piping of long transcripts

See docs/testing.md for the full report.

10+ Demo Scripts

Real-world shell scripts showcasing ohr's capabilities.

ohr only

• subtitle — generate SRT/VTT subtitles with --save
• batch-transcribe — transcribe all audio files in a directory
• audio-grep — grep-like search with timestamps and JSON output
• voice-search — search inside audio files by spoken content
• dictate — speak into a text file via live microphone
• live-caption — real-time captions in the terminal
• whisper-compat — drop-in replacement for OpenAI Whisper CLI

ohr + apfel

• minutes — meeting recording to structured meeting minutes
• action-items — extract to-dos and commitments from meetings
• translate-audio — transcribe then translate to any language
• podcast-chapters — generate timestamped chapter markers
• voice-note — record, transcribe, and optionally summarize

Quick examples

demo/subtitle lecture.m4a --save
demo/audio-grep "budget" meetings/*.m4a
demo/minutes standup.m4a -o markdown > standup.md
demo/batch-transcribe ~/recordings/ -o srt
demo/whisper-compat audio.m4a --output_format srt

QA + Homebrew + Integration Tests

QA Fixes

• Fixed stdin piping — binary audio was corrupted by readLine, now uses raw FileHandle.readDataToEndOfFile with magic byte format detection
• All CLI paths verified: file, stdin, srt, vtt, json, timestamps, language, errors
• All server endpoints verified: health, models, transcription (5 formats), auth, CORS, stubs

Integration Tests (42 tests, all green)

• cli_e2e_test.py — help, version, exit codes, file transcription, output formats, stdin
• server_test.py — health, models, transcription (json/verbose_json/text/srt/vtt), error handling, stubs
• security_test.py — token auth, origin validation, CORS preflight

Homebrew

brew tap Arthur-Ficial/tap
brew install Arthur-Ficial/tap/ohr

Test Coverage

• 109 unit tests (Swift, pure logic)
• 42 integration tests (Python, end-to-end)
• Total: 151 tests

Binary

• ohr-0.1.3-arm64-macos.tar.gz — Apple Silicon macOS binary

Full Implementation — CLI + Server + Transcription Engine

ohr is now a complete, production-grade tool matching apfel's architecture.

CLI Tool

ohr meeting.m4a # Transcribe → plain text
ohr -o srt lecture.wav # Generate SRT subtitles
ohr -o vtt interview.m4a # Generate VTT subtitles
ohr -o json recording.mp3 # JSON with segments and timestamps
ohr --timestamps meeting.m4a # Plain text with timestamps
ohr meeting.m4a | apfel "summarize" # Pipe to apfel
ohr --listen # Live microphone transcription
cat audio.wav | ohr # Transcribe from stdin

OpenAI-Compatible HTTP Server

ohr --serve # Start server on :11434
curl -X POST http://localhost:11434/v1/audio/transcriptions \
 -F file=@meeting.m4a -F model=apple-speechanalyzer

Features

• 6 test suites, 109 tests (all green, TDD red-to-green)
• 4 output formats: plain text, JSON, SRT subtitles, VTT subtitles
• OpenAI API compatible: POST /v1/audio/transcriptions, GET /v1/models, GET /health
• Security: origin validation, Bearer token auth, CORS control
• 30 languages supported via Apple's on-device SpeechAnalyzer
• 100% on-device — no cloud, no API keys, no network for inference
• Zero dependencies beyond Hummingbird (HTTP server) and macOS 26 SDK

Architecture (same as apfel)

• OhrCore — pure-logic library (testable without Speech framework)
• ohr — executable (CLI + HTTP server)
• ohr-tests — pure Swift test runner (no XCTest needed)

OhrCore Library — Foundation Complete

First release of ohr: on-device speech-to-text from the command line.

What's in this release

OhrCore pure-logic library with 6 test suites, 109 tests (all green):

• OhrError — Speech-to-text error classification with OpenAI API error mapping
• AudioFormat — Audio format detection from file extensions and MIME types (m4a, wav, mp3, caf, aiff, flac, mp4)
• SubtitleFormatter — SRT and VTT subtitle generation with precise timestamp formatting
• OpenAIModels — Transcription API types (TranscriptionResponse, VerboseTranscriptionResponse, segments)
• TranscriptionValidator — Request parameter validation (format, temperature, response format)
• OriginValidator — Localhost CSRF protection and Bearer token authentication

Project scaffolding

• Package.swift with 3 targets: OhrCore (library), ohr (executable), ohr-tests (test runner)
• Makefile with version bumping, build, install, release asset packaging
• Pure Swift test runner (no XCTest needed, Command Line Tools only)
• Same architecture as apfel

What's next

CLI implementation, SpeechAnalyzer integration, and OpenAI-compatible HTTP server.