speaker-detector 0.2.3

A CLI + Web tool for speaker enrollment and identification using SpeechBrain.

08/08: a lot of improvements to the speaker-detector-client which resulted in a few changes to this backend.

23/07/2025 - Lara Whybrow, Creator - it has a few bugs that need fixing, but I ma determining if it is data related or software related. Feel free to clone from Github and help with bug fixes.

speaker-detector 🎙️

A lightweight CLI tool for speaker enrollment and voice identification, powered by SpeechBrain.

🔧 Features

• ✅ Enroll speakers from .wav audio
• 🕵️ Identify speakers from audio samples
• 🧠 ECAPA-TDNN embedding-based matching
• 🎛️ Simple, fast command-line interface
• 📁 Clean file storage in ~/.speaker-detector/
• 🔊 Optional --verbose mode for debugging

Web UI note: The web client uses a guided-only enrollment flow (multiple short recordings). Quick enroll with a single clip has been removed to ensure model accuracy.

📦 Installation

pip install speaker-detector


When installing packages with a stale requirement file you might need to use:  pip install --break-system-packages soundfile to install on WSL Ubuntu environment.

Run this version with -m module flag if you are having issues with running server.py:
python3 -m speaker_detector.server

🚀 Example Usage

🎙️ Enroll a speaker:

speaker-detector record --enroll Lara

🕵️ Identify a speaker:

speaker-detector record --test

📋 List enrolled speakers:

speaker-detector list

🗂️ Project Structure

~/.speaker-detector/enrollments/ Saved .pt voice embeddings ~/.speaker-detector/recordings/ CLI-recorded .wav audio files

🧹 Clean vs Verbose Mode By default, warnings from speechbrain, torch, etc. are hidden for a clean CLI experience. To enable full logs & deprecation warnings:

speaker-detector --verbose identify samples/test_sample.wav

🛠 Requirements Python 3.8+ torch speechbrain numpy soundfile onnxruntime

Step	Command	When / Purpose	Output
1. Export ECAPA Model to ONNX	speaker-detector export-model --pt models/embedding_model.ckpt --out ecapa_model.onnx	Run once unless model changes	ecapa_model.onnx
2. Enroll Speaker	speaker-detector enroll <speaker_id> <audio_path> Example: speaker-detector enroll Lara samples/lara1.wav	Run per new speaker	Individual .pt files (e.g., Lara.pt)
3. Combine Embeddings	speaker-detector combine --folder data/embeddings/ --out data/enrolled_speakers.pt	After enrolling speakers	enrolled_speakers.pt
4. Export Speakers to JSON	speaker-detector export-speaker-json --pt data/enrolled_speakers.pt --out public/speakers.json	For frontend use	speakers.json
5. Identify Speaker	speaker-detector identify samples/test_sample.wav	Identify speaker from audio	Console output: name + score
6. List Enrolled Speakers	speaker-detector list-speakers	Show all enrolled speakers	Console output: list of IDs
Verbose Mode (optional)	Add --verbose to any command: speaker-detector --verbose identify samples/test_sample.wav	Show warnings, detailed logs	Developer debug info

NB: When pushing to Github, do not include any .identifier files.

You can manually clean up stale embeddings that don’t match any existing speaker folder with a quick script:

Run inside your project root

cd storage/embeddings for f in *.pt; do speaker="${f%.pt}" if [ ! -d "../speakers/$speaker" ]; then echo "Deleting stale embedding: $f" rm "$f" fi done

HTTP API: Online & Detection State

This backend exposes simple endpoints to let a client know when the server is reachable and when live detection is ready to be polled.

Online (one-shot SSE)

• Path: GET /api/online
• Headers:
  • Content-Type: text/event-stream
  • Cache-Control: no-cache
  • Connection: keep-alive
  • Access-Control-Allow-Origin: http://localhost:5173 (override with env CLIENT_ORIGIN)
• Behavior: immediately emits a single event and closes the stream.

Example event:

event: online
data: 1

This removes the need for heartbeat polling: as soon as the client connects, it can mark the backend as reachable.

Detection State (SSE)

• Path: GET /api/detection-state
• Emits an immediate state and then re-emits on changes; includes keep-alives.
• Event name: detection
• Data: running | stopped

Example stream excerpts:

event: detection
data: stopped

: keep-alive

event: detection
data: running

Clients can start polling /api/active-speaker only when the state is running, and pause when stopped.

Active Speaker (readiness semantics)

• Path: GET /api/active-speaker
• Responses:
  • When listening mode is OFF: 200 { "status": "disabled", "speaker": null, "confidence": null, "is_speaking": false }
  • When mode is ON but engine not yet ready (e.g., mic unavailable or loop not running): 200 { "status": "pending", ... }
  • When running and healthy: 200 with the usual payload including speaker, confidence, is_speaking, status: "listening", and optional suggested.

These semantics avoid red 503s in DevTools while still making state transitions explicit for the client.

Quick Examples

Curl (SSE streams)

# One-shot online event
curl -N -H 'Accept: text/event-stream' http://127.0.0.1:9000/api/online

# Detection state stream (emits running|stopped)
curl -N -H 'Accept: text/event-stream' http://127.0.0.1:9000/api/detection-state

Browser client (minimal)

// Reachability: mark backend online as soon as server is up
const online = new EventSource('http://127.0.0.1:9000/api/online');
online.addEventListener('online', () => {
  console.log('Backend online');
  online.close(); // one-shot
});

// Detection state: start/stop polling active speaker
let pollTimer = null;
function startPolling() {
  if (pollTimer) return;
  pollTimer = setInterval(async () => {
    try {
      const r = await fetch('http://127.0.0.1:9000/api/active-speaker');
      const j = await r.json();
      if (j.status === 'disabled' || j.status === 'pending') return; // wait
      console.log('Active:', j);
    } catch (e) {
      console.warn('poll failed', e);
    }
  }, 500);
}
function stopPolling() { clearInterval(pollTimer); pollTimer = null; }

const detect = new EventSource('http://127.0.0.1:9000/api/detection-state');
detect.addEventListener('detection', (ev) => {
  const state = (ev.data || '').trim();
  if (state === 'running') startPolling(); else stopPolling();
});