Skip to main content

​
Overview

The NativeVad class provides Voice Activity Detection (VAD) capabilities for analyzing audio frames to detect speech activity. It’s useful for building voice-activated features, speech detection systems, and audio bots that need to identify when someone is speaking. VAD is commonly used in conversational AI applications to determine when a user has finished speaking, enabling natural turn-taking in voice interactions.

​
Creation

Create a NativeVad instance using the static method Daily.create_native_vad(): 
from daily import Daily

# Initialize Daily SDK
Daily.init()

# Create a VAD instance
vad = Daily.create_native_vad(
    reset_period_ms=500,
    sample_rate=16000,
    channels=1
)

​
Parameters

​
reset_period_ms
int
default:"500"
The period in milliseconds after which the VAD state resets. This determines how long the VAD maintains its internal state before resetting.
​
sample_rate
int
default:"16000"
The audio sample rate in Hz. Must match the sample rate of the audio frames being analyzed. Common values are 8000, 16000, 24000, or 48000.
​
channels
int
default:"1"
The number of audio channels. Use 1 for mono, 2 for stereo.

​
Properties

​
rest_period_ms
int
The configured reset period in milliseconds.
print(f"Reset period: {vad.rest_period_ms}ms")
​
sample_rate
int
The configured audio sample rate in Hz.
print(f"Sample rate: {vad.sample_rate}Hz")
​
channels
int
The configured number of audio channels.
print(f"Channels: {vad.channels}")

​
Methods

​
analyze_frames()

def analyze_frames(self, frame: bytes) -> float
Analyzes audio frames to detect voice activity.

​
Parameters

​
frame
bytes
required
Raw audio frame data as bytes. The frame should match the configured sample rate and number of channels.

​
Returns

​
confidence
float
A confidence score between 0.0 and 1.0 indicating the likelihood of speech being present in the audio frame. Higher values indicate greater confidence that speech is detected.
  • 0.0 - 0.5: Likely silence or background noise
  • 0.5 - 0.8: Possible speech or ambiguous audio
  • 0.8 - 1.0: High confidence speech detected

​
Usage Example

Here’s a complete example demonstrating speech detection using NativeVad: 
from daily import Daily, CallClient
import time
from enum import Enum

SAMPLE_RATE = 16000
CHANNELS = 1
SPEECH_THRESHOLD = 0.90
RESET_PERIOD_MS = 2000

class SpeechStatus(Enum):
    SPEAKING = 1
    NOT_SPEAKING = 2

class SpeechDetector:
    def __init__(self):
        # Create VAD instance
        self.vad = Daily.create_native_vad(
            reset_period_ms=RESET_PERIOD_MS,
            sample_rate=SAMPLE_RATE,
            channels=CHANNELS
        )
        self.status = SpeechStatus.NOT_SPEAKING
        
    def analyze(self, audio_buffer):
        # Get confidence score from VAD
        confidence = self.vad.analyze_frames(audio_buffer)
        
        # Determine speech status based on threshold
        if confidence > SPEECH_THRESHOLD:
            self.status = SpeechStatus.SPEAKING
            print(f"SPEAKING: {confidence:.2f}")
        else:
            self.status = SpeechStatus.NOT_SPEAKING
            print(f"NOT SPEAKING: {confidence:.2f}")
        
        return self.status

# Initialize Daily SDK
Daily.init()

# Create speech detector
detector = SpeechDetector()

# Create a speaker device to receive audio
speaker = Daily.create_speaker_device(
    "my-speaker",
    sample_rate=SAMPLE_RATE,
    channels=CHANNELS
)
Daily.select_speaker_device("my-speaker")

# Create and join call
client = CallClient()
client.join("https://your-domain.daily.co/room")

# Process audio in a loop
try:
    while True:
        # Read 10ms worth of audio frames
        buffer = speaker.read_frames(int(SAMPLE_RATE / 100))
        if len(buffer) > 0:
            detector.analyze(buffer)
        time.sleep(0.01)
except KeyboardInterrupt:
    client.leave()

​
Advanced Example

For a more sophisticated implementation with configurable thresholds and state management, see the native_vad.py demo in the Daily Python SDK repository. The demo includes:
  • Configurable speech and silence thresholds
  • Time-based state transitions for more accurate detection
  • Command-line arguments for tuning VAD parameters
  • Integration with Daily’s virtual speaker device

​
Common Use Cases

Use VAD to detect when a user has finished speaking, allowing your bot to respond at natural breaks in conversation:
def on_silence_detected():
    if user_was_speaking:
        # User finished speaking, bot can respond now
        generate_and_play_response()
Start and stop recording based on voice activity to save storage and processing:
if speech_confidence > 0.9 and not recording:
    start_recording()
elif speech_confidence < 0.3 and recording:
    stop_recording()
Only send audio to transcription services when speech is detected:
if vad.analyze_frames(buffer) > SPEECH_THRESHOLD:
    send_to_transcription_service(buffer)

​
Tips

Threshold Tuning: Start with a confidence threshold of 0.90 for detecting speech. Adjust based on your environment - use higher values (0.95+) for noisy environments or lower values (0.70-0.85) for quiet environments.
Sample Rate Matching: Ensure the VAD’s sample rate matches your audio source. Mismatched sample rates will produce inaccurate results.
The NativeVad analyzes each frame independently. For production applications, combine VAD confidence scores with time-based thresholds to avoid rapid state changes from brief noise or pauses.

Build docs developers (and LLMs) love

Get started for freeTalk to us