Yobe BioPSI SDK

A Python SDK for real-time speaker verification, ambient calibration, and transcription. Processes 1-second audio chunks on-device. No GPU required.

Python 3.8+  ·  macOS  ·  Windows  ·  16 kHz mono audio

The SDK provides five capabilities:

• Ambient calibration. Set a per-user energy threshold from background noise.
• Speaker enrollment. Train a voice profile from streamed audio chunks.
• Speaker verification. Confirm whether a chunk contains speech from the enrolled user (1:1).
• Transcription. Whisper-based speech-to-text, emitted on detected pauses.
• Real-time streaming. 1-second chunked processing.

Installation

1. Extract the SDK into your project directory.

2. Install Python dependencies.

pip install numpy pyaudio

3. Place the binary. Copy the audio_processor binary (delivered with your license email) into ./dist/. On Windows this is audio_processor.exe.

4. Add your license file.

echo "YOUR-LICENSE-KEY" > dist/license.txt

The final ./dist/ directory should contain audio_processor and license.txt.

Hello world

Verify the SDK loads and the license is valid:

from yobe_client import YobeClient

with YobeClient('./dist/audio_processor') as client:
    print("SDK initialized.")

If this runs without an exception, the binary started and the license validated.

Core workflow

The integration has three stages: calibrate, enroll, verify. The example below uses pyaudio to capture from the default microphone.

from yobe_client import YobeClient
import numpy as np
import pyaudio

def record_chunk(stream, samples=16000):
    raw = stream.read(samples, exception_on_overflow=False)
    return np.frombuffer(raw, dtype=np.int16).astype(np.float32)

pa = pyaudio.PyAudio()
stream = pa.open(format=pyaudio.paInt16, channels=1,
                 rate=16000, input=True, frames_per_buffer=16000)

with YobeClient('./dist/audio_processor') as client:

    # 1. Calibrate to the environment
    ambient = np.concatenate([record_chunk(stream) for _ in range(3)])
    cal = client.calibrate_chunk("alice", ambient, duration_seconds=3)
    print(f"Threshold: {cal.threshold}")

    # 2. Enroll: stream chunks until train_chunk returns True
    while True:
        done = client.train_chunk(
            "alice",
            record_chunk(stream),
            on_progress=lambda p: print(f"Training: {p:.1f}%"),
            on_complete=lambda: print("Done."),
        )
        if done: break

    # 3. Verify + transcribe in real time
    while True:
        result = client.test_chunk("alice", record_chunk(stream))

        if result.has_auth and result.authenticated:
            authorized = result.authorized_audio  # non-enrolled speech zeroed

        if result.has_transcription:
            print(result.text)

API

YobeClient(binary_path)

Constructs a client and starts (or attaches to) the background service. Auto-detects the .exe extension on Windows. Use as a context manager for automatic cleanup.

client = YobeClient(binary_path='./dist/audio_processor')
# or
with YobeClient('./dist/audio_processor') as client:
    ...

client.calibrate_chunk(username, audio_chunk, duration_seconds=3)

Establishes a per-user energy threshold from ambient noise. The threshold is persisted automatically.

Returns an object with a threshold attribute (float).

client.train_chunk(username, audio_chunk, on_progress=None, on_complete=None)

Streams a 1-second chunk into the user's voice profile. Call repeatedly until it returns True.

Returns True once sufficient training data has been collected.

client.test_chunk(username, audio_chunk)

Runs verification on the chunk and, when a pause is detected, transcription.

Returns a result object — see Result objects.

client.reset(username, mode)

Clears session state. mode is "train", "test", or "both".

client.close()

Gracefully shuts down the client. Called automatically by the context manager.

Result objects

Always check the has_* flags before reading the corresponding value.

Audio requirements

Input must match exactly. Mismatches raise InvalidAudioFormatError.

Errors

All SDK exceptions inherit from YobeError.

How it works

The SDK runs a persistent local service. The first call in any process launches it; subsequent scripts attach over loopback TCP.

Your Python application
      ↓ imports
YobeClient  (yobe_client.py)
      ↓ subprocess + TCP
audio_processor  (native binary)
      ↓ processes
Speaker recognition  +  Whisper transcription

• First call spawns the audio_processor service.
• Later scripts detect the running service and attach — startup is near-instant.
• To stop the service manually: python examples/shutdown_service.py.

Project structure

System requirements

Bundled examples

Run in this order from a fresh checkout:

cd examples

python voice_calibration.py     # 1. calibrate threshold (run first)
python voice_enrollment.py      # 2. enroll a user voice profile
python voice_verification.py    # 3. verify in real time

python shutdown_service.py      # stop background service when done

Troubleshooting

License file not found

LicenseError: License file not found at: ./dist/license.txt

Create license.txt alongside the audio_processor binary and paste your key into it.

Binary not found

FileNotFoundError: Binary not found at: ./dist/audio_processor

Confirm dist/ is in your project root with the correct binary for your OS (.exe on Windows).

Service connection error

ServiceConnectionError: Failed to connect to existing service

Stop any stale service and let the next call respawn it:

python examples/shutdown_service.py

Invalid audio format

InvalidAudioFormatError: Invalid audio format

Audio must be 16 kHz, mono, Float32 or Int16, and exactly 16,000 samples per chunk.

License

SDK Evaluation Agreement (non-commercial). Copyright © 2026 Yobe Inc.

License grant

The SDK is provided as a compiled library for evaluation and testing purposes only. Yobe Inc. grants a personal, non-exclusive, non-transferable license to integrate the SDK into development projects for internal exploration.

Prohibited uses

You agree not to:

• Commercialize — no production use, commercial product, or third-party service offering.
• Reverse engineer — no decompiling, disassembly, or source-code derivation.
• Distribute — no sharing, renting, leasing, or sublicensing of the binaries.
• Bypass — no circumvention of license-key validation or technical restrictions.

License keys

Operation requires a valid license key issued by Yobe Inc. Keys are tied to the evaluation period and may be revoked or expired at Yobe's discretion.

Ownership

The SDK and all intellectual property rights remain the sole property of Yobe Inc.

Limitation of liability

The SDK is provided "AS IS." To the maximum extent permitted by law, Yobe Inc. shall not be liable for any damages arising out of the use or inability to use the SDK.

Support

Technical support, integration help, or licensing questions: contact.us@yobeinc.com

For implementation patterns, see the examples/ directory.