ohr-speaker 🔊

On-device speech-to-text with speaker diarization, powered by Apple Intelligence & CAM++

中文版 | English

ohr-speaker is an enhanced fork of ohr that integrates three speaker diarization engines — including CAM++ (阿里达摩院) and FluidAudio — on top of the original Apple SpeechAnalyzer transcription engine.

100% on-device. No cloud. No API keys. Your data never leaves your machine.

---

Features

• 🎤 Apple Intelligence Transcription — Powered by macOS SpeechAnalyzer, millisecond latency
• 🗣️ Three Diarization Engines — CAM++, Offline VBx (WeSpeaker), or Sortformer (--engine flag)
• 🇨🇳 Chinese-Optimized — CAM++ model trained on 200k Chinese speakers, ~15MB model
• 📝 Multiple Output Formats — Plain text, JSON, SRT subtitles, VTT subtitles with speaker labels
• 🎙️ Live Microphone Transcription — Real-time --listen mode
• 🖥️ OpenAI-Compatible Server — --serve mode, compatible with OpenAI Whisper API
• 📦 Multi-format Support — m4a, wav, mp3, mp4, caf, aiff, flac
• 🔒 Privacy First — All processing is local, no data uploaded

Quick Start

Download a pre-built binary from Releases:

Binary	Default Engine	Model Size
ohr-cam	CAM++ (default)	~15 MB
ohr-speaker	Offline VBx (WeSpeaker)	~700 MB

# CAM++ engine (~15 MB model, auto-downloaded on first use)
ohr-cam --speakers meeting.wav

# Or offline VBx engine (legacy, ~700 MB model)
ohr-speaker --speakers meeting.wav

Build from Source

git clone https://github.com/yanhuicsdn/ohr-speaker.git
cd ohr-speaker

# Using Xcode-beta toolchain (macOS 27+)
DEVELOPER_DIR="/Applications/Xcode-beta.app/Contents/Developer" swift build -c release

# Using Xcode stable (macOS 26+)
swift build -c release

# Copy the binary to your PATH
cp .build/release/ohr /usr/local/bin/ohr-speaker

Prerequisites

• macOS 26+ (Apple Silicon)
• Xcode 26.6+ or Xcode-beta 27+

Usage

Basic Transcription

ohr-speaker audio.wav

With Speaker Diarization (CAM++ default)

ohr-speaker --speakers audio.wav

Sample output:

【S1】
There is usually a platform management layer with a user center...

【S2】
Give them the full version so they can use it...

【S1】
Yes, provincial business layers have isolation requirements for permissions...

Select a Different Diarization Engine

# CAM++ (default) — 阿里达摩院, 200k Chinese speakers, ~15MB model
ohr-speaker --speakers --engine campplus audio.wav

# Offline VBx (WeSpeaker) — ~700MB model, PLDA scoring
ohr-speaker --speakers --engine offlineVbx audio.wav

# Sortformer — ~200MB model, ≤4 speakers, CoreML ANE accelerated
ohr-speaker --speakers --engine sortformer audio.wav

Output as SRT Subtitles

ohr-speaker --speakers -o srt audio.wav > subtitles.srt

Output as JSON

ohr-speaker --speakers -o json audio.wav

Live Microphone Transcription

ohr-speaker --listen --speakers

Start OpenAI-Compatible Server

ohr-speaker --serve --port 11434

Then call from any OpenAI Whisper client:

curl http://localhost:11434/v1/audio/transcriptions \
  -F "file=@audio.wav" \
  -F "model=whisper-1" \
  -F "diarize=true"

Complete Options

USAGE:
  ohr <file>                   Transcribe an audio file
  ohr -o srt <file>            Transcribe to SRT subtitles
  ohr -o vtt <file>            Transcribe to VTT subtitles
  ohr -o json <file>           Transcribe to JSON with segments
  ohr --listen                 Live microphone transcription
  ohr --serve                  Start OpenAI-compatible HTTP server
  cat audio.wav | ohr          Transcribe from stdin

OPTIONS:
  -o, --output <format>        Output: plain (default), json, srt, vtt
  --json                       Shorthand for -o json
  --srt                        Shorthand for -o srt
  --vtt                        Shorthand for -o vtt
  --timestamps                 Show timestamps in plain text output
  --speakers                   Enable speaker diarization
  --engine <name>              Diarization engine: campplus, offlineVbx, sortformer
  -l, --language <code>        Language code (e.g. en-US, de-DE)
  -q, --quiet                  Suppress headers and chrome
  --no-color                   Disable ANSI colors

SERVER OPTIONS:
  --serve                      Start HTTP server
  --port <n>                   Server port (default: 11434, env: OHR_PORT)
  --host <addr>                Bind address (default: 127.0.0.1, env: OHR_HOST)
  --cors                       Enable CORS headers
  --allowed-origins <list>     Comma-separated allowed origins
  --no-origin-check            Disable origin validation
  --token <secret>             Require Bearer token (env: OHR_TOKEN)
  --token-auto                 Generate random token on startup
  --public-health              /health without auth on non-loopback
  --footgun                    Disable all protections (DANGEROUS)
  --max-concurrent <n>         Max concurrent requests (default: 5)
  --debug                      Enable /v1/logs endpoints

INFO:
  -h, --help                   Show this help
  -v, --version                Show version
  --release                    Show detailed build info
  --model-info                 Show model capabilities

Diarization Engine Comparison

Engine	Model	Parameters	Size	Speakers	M4 Accel.	Language Bias
CAM++ ⭐	阿里达摩院 CAM++	7.2M	~15 MB	unlimited	GPU	Chinese ✅
Offline VBx	WeSpeaker + PLDA	~25M	~700 MB	unlimited	ANE ✅	Multi
Sortformer	FluidAudio Sortformer	—	~200 MB	≤4	ANE ✅	Multi
LS‑EEND ❌	dihard3 EEND	—	~44 MB	unlimited	CoreML	English only

Benchmark (2-speaker Chinese podcast, 17 min):

Engine	Segments	S1	S2	First-run download
CAM++ ⭐	22	11	11	~5 sec
Offline VBx	26	13	13	~2 min
Sortformer	26	13	13	~1 min

CAM++ is the recommended engine for Chinese audio — smallest model, fastest download, specifically trained on 200k Chinese speakers by Alibaba DAMO Academy.

Performance

Audio Duration	Transcription Only	Additional Diarization
1 min	~1s	~1s
10 min	~3s	~2s
31 min	~10s	~5s
1 hour	~20s	~10s

How It Works

1. Transcription: Uses Apple's SpeechAnalyzer + SpeechTranscriber framework to convert audio into timestamped text segments
2. Diarization: One of three engines:
   • CAM++: FluidAudio segmentation → independent CAM++ embedding extractor (192-d) → cosine agglomerative clustering
   • Offline VBx: FluidAudio's full pipeline (Seg-3 + WeSpeaker + PLDA + VBx)
   • Sortformer: FluidAudio's end-to-end Sortformer (≤4 speakers, CoreML ANE)
3. Alignment: Uses a time-overlap-maximization algorithm to align transcription segments with speaker segments

Environment Variables

Variable	Description
OHR_PORT	Server port (default: 11434)
OHR_HOST	Server bind address (default: 127.0.0.1)
OHR_TOKEN	Bearer token for server auth
OHR_LANGUAGE	Default language code
REGISTRY_URL	Model download mirror (e.g. https://hf-mirror.com)
NO_COLOR	Disable ANSI colors

Differences from Upstream ohr

Feature	ohr	ohr-speaker
Speaker Diarization	❌	✅ --speakers
Transcription Engine	Apple SpeechAnalyzer	Apple SpeechAnalyzer
Diarization Engine	—	CAM++ (default) / Offline VBx / Sortformer
Model Download	0 MB	~15 MB (CAM++) or ~700 MB (VBx)
Engine Selection	—	--engine flag
Output Formats	plain/json/srt/vtt	plain/json/srt/vtt + speaker labels
Server Mode	✅	✅ (supports diarize parameter)

Model Cache Locations

Engine	Cache Path	Size
CAM++	~/Library/Application Support/ohr-campplus/	~15 MB
FluidAudio	~/Library/Application Support/FluidAudio/Models/	~700 MB

Credits

• Arthur-Ficial/ohr — The original ohr project
• FluidInference/FluidAudio — FluidAudio diarization engine
• 3D-Speaker / FunASR — CAM++ speaker embedding by Alibaba DAMO Academy
• Apple Speech Framework — macOS built-in speech recognition capabilities

License

MIT