Skip to content

Bananaflow MCP is live.

Read the docs

Silero VAD on 8 kHz Indian Phone Audio

Jaseem Abbas, founder of Bananaflow, in glasses and a dark top against a pale background.Jaseem Abbas

Published
Cover for "Silero VAD on 8 kHz Indian Phone Audio": a white line diagram of a voice detector picking out the caller: three glass panels stacked on the left, a car, a person and a speaker, each send a thin wave toward one lit node; the car's and the speaker's waves end at the node as faint dashed lines, and only the person's passes through as a bold arrow to a panel with a sound wave labelled Caller, under a pill reading Which is the caller?, drawn over a soft, out-of-focus salvia in muted sky blue, its slender petals blurred into veils, with an aqua haze at one end and a dusty rose glow at the other.

Silero VAD is an open-source voice activity detector: a small model that listens to audio in 32 millisecond chunks and gives each one a probability of being speech. This guide is for engineers who run it on phone audio. It covers the 8 kHz sample rate, the chunk sizes, the settings that matter, how Indian street and home noise shows up, and how to tune and test it. For the plain definition, see our glossary entry on voice activity detection.

Here is the problem in one call. A clinic's agent in Pune calls a patient, Sunita, who answers from a balcony. A scooter passes. A TV serial plays in the next room. Her mother says "Sunita, kaun hai?" from the kitchen. Then Sunita says "haan, bolo". Four sounds, and only one of them is the caller speaking to the agent. A detector with the wrong settings treats all four as the caller, and the agent keeps stopping. Sunita, the clinic and the street are made up for this example.

What is Silero VAD, and how does it work?

Silero VAD is a pre-trained voice activity detector that you run yourself. Its repository describes it as a "pre-trained enterprise-grade Voice Activity Detector" and says it was trained on corpora of over 6000 languages (snakers4/silero-vad README, read 16 November 2026). Its JIT model is about two megabytes, and one audio chunk of 30 milliseconds or more takes less than a millisecond on a single CPU thread, by the repository's own figures.

It works in three steps:

  1. You give it a chunk of audio of a fixed size.
  2. It returns a number between 0 and 1, the probability that the chunk holds speech.
  3. Your code compares that number to a threshold and decides.

The model does not decide where a sentence begins or ends. That is the job of the code around it, which uses the threshold and the timing settings you choose. The repository's quality notes say the same thing: modern detectors output a probability, and the threshold should be tailored to the result you want (Silero VAD quality metrics, read 16 November 2026).

VAD is the bottom layer of turn-taking. The layers above it decide when a caller has finished and when to let them cut in. We cover them in our guides to endpointing and turn detection and to barge-in on voice agents.

Does Silero VAD work on 8 kHz phone audio?

Yes. Silero VAD supports 8000 Hz and 16000 Hz, and 8000 Hz is the sample rate of a normal phone call. The repository lists 8000 Hz and 16000 Hz as its supported sampling rates (README, key features, read 16 November 2026).

Why 8 kHz is the phone rate: the G.711 codec, the standard voice codec on phone lines, is an ITU-T recommendation titled "Pulse code modulation (PCM) of voice frequencies", approved in 1988 and still in force on the day we read the record (ITU-T G.711). The IETF profile for sending it over RTP says each sample is eight bits, and the payload runs at 8,000 samples a second (RFC 3551, read 16 November 2026). So a call that reaches your agent over a normal SIP trunk carries 8,000 samples every second.

A sampling rate of 8,000 a second can carry sound only up to 4 kHz. That is enough for speech to be understood, and not enough for the crisp "s" and "f" sounds that 16 kHz audio keeps. A detector working on phone audio sees a duller picture than one working on a laptop microphone, and the settings that suit a laptop may not suit a phone line.

The path of phone audio to the detector: 8-bit samples at 8,000 a second come off the carrier line, are cut into 256-sample chunks, and each chunk gets a speech probability that your code compares with a threshold.

Keep the audio at its native rate where you can. The model takes 8 kHz directly, so you do not need to convert phone audio up to 16 kHz before it reaches the detector. If your pipeline already runs at 16 kHz for another reason, use the 16 kHz setting and its own chunk size. Never mix rates inside one stream: the repository's code resets the model's state when the rate changes.

What chunk size does Silero VAD need?

Silero VAD needs exactly 256 samples per chunk at 8000 Hz and exactly 512 samples at 16000 Hz. Both are 32 milliseconds. The model code raises an error for any other size, saying that 256 is supported for 8000 and 512 for 16000.

  • Sample rate: 8000 Hz (phone audio)

    Samples per chunk
    256
    Length of a chunk
    32 ms
    Context the model adds
    32 samples
  • Sample rate: 16000 Hz

    Samples per chunk
    512
    Length of a chunk
    32 ms
    Context the model adds
    64 samples

On a live call, this means your agent hears the line in 32 millisecond steps. The packets a carrier sends are not tied to that size, so your code needs a small buffer that collects samples until it has a full chunk. A detector can never react faster than one chunk, so about 32 milliseconds is the floor on its reaction time, before any padding or minimum length.

What do the Silero VAD settings do?

Four settings do most of the work: the threshold, the minimum speech length, the minimum silence length and the speech padding. Here are the parameters of get_speech_timestamps in the repository, with their defaults and what they do, read in src/silero_vad/utils_vad.py on 16 November 2026.

  • Setting: threshold

    Default
    0.5
    What it does
    Chunks with a speech probability above this count as speech
    Too low
    Noise counts as speech
    Too high
    Soft voices are missed
  • Setting: neg_threshold

    Default
    threshold minus 0.15
    What it does
    Once in speech, chunks below this count as not speech
    Too low
    The segment hangs on in noise
    Too high
    The segment breaks in the middle of a word
  • Setting: min_speech_duration_ms

    Default
    250
    What it does
    Speech shorter than this is thrown out
    Too low
    Clicks and coughs get through
    Too high
    A short "haan" or "ji" is dropped
  • Setting: min_silence_duration_ms

    Default
    100
    What it does
    How long to wait at the end of speech before closing the segment
    Too low
    The caller is cut off at a pause
    Too high
    Dead air after the caller finishes
  • Setting: speech_pad_ms

    Default
    30
    What it does
    Padding added to each side of a segment
    Too low
    Word edges are clipped
    Too high
    A little noise is dragged in

Two notes on this table.

  • The repository says "it is better to tune this parameter for each dataset separately, but 'lazy' 0.5 is pretty good for most datasets". That is the best starting point and a poor stopping point.
  • The 100 millisecond minimum silence is a segmenter default, not a turn decision. Deciding that a caller has finished is the job of endpointing, which waits longer. See endpointing and turn detection.
Four settings on one timeline of probabilities. The threshold and the exit threshold set when speech starts and stops, the minimum speech length removes a short spike, and the padding widens the segment on both sides.

Here is the official starting code, from the repository's README, with only the audio rate changed to match a phone line. The call and its argument names are the repository's own; the sampling_rate value is the only edit.

from silero_vad import load_silero_vad, read_audio, get_speech_timestamps

model = load_silero_vad()
wav = read_audio("path_to_audio_file", sampling_rate=8000)
speech_timestamps = get_speech_timestamps(
    wav,
    model,
    sampling_rate=8000,
    threshold=0.5,
    min_speech_duration_ms=250,
    min_silence_duration_ms=100,
    speech_pad_ms=30,
    return_seconds=True,
)

For a live call you stream chunks instead of reading a file. The repository ships a VADIterator class for this. It takes the model, threshold, sampling_rate (8000 or 16000), min_silence_duration_ms, speech_pad_ms and neg_threshold, and you call it with one chunk at a time. Check the examples folder of the repository for the current streaming code before you copy anything, because the library changes between releases.

How do Indian traffic, TV and family noise show up in the detector?

They show up in four different ways, and each one needs a different fix. The table maps each sound to what the detector does and what to try. These patterns come from how detectors work in general, so confirm each one on your own recordings.

  • Sound: Traffic and horns

    What the detector may do
    Short spikes of high probability that start false speech
    What to try
    Raise the threshold a little, and lengthen the minimum speech
  • Sound: A fan, a cooler or an AC

    What the detector may do
    A steady low probability that rarely crosses the line
    What to try
    Usually nothing. If it does cross, raise the threshold
  • Sound: A TV or radio in the room

    What the detector may do
    It is human speech, so the model may call it speech
    What to try
    A higher threshold helps only a little. Keep agent steps short on these calls
  • Sound: A second voice in the room

    What the detector may do
    Real speech that is not your caller
    What to try
    VAD cannot tell speakers apart. Use words and the step design to decide
  • Sound: A soft or distant voice

    What the detector may do
    Probability stays low and speech is missed
    What to try
    Lower the threshold, and keep the minimum speech short
  • Sound: The agent's own voice on speakerphone

    What the detector may do
    The agent's voice comes back and counts as the caller
    What to try
    Test on speakerphone, and raise the threshold if it happens

The second voice is the hard case. Silero VAD says whether there is speech, not whose speech it is. When Sunita's mother says "kaun hai?", the detector hears speech. Something above it has to decide that the words are not for the agent. That is one reason barge-in looks at words as well as sound. Our guide to barge-in on voice agents covers that decision.

Network effects matter too. Some carriers send nothing during silence. The IETF's audio profile has a section on silence suppression for this (RFC 3551, section 4.1), and ITU-T G.729 Annex B, approved in 1996, is a silence compression scheme for a different phone codec (ITU-T G.729 Annex B). If your carrier uses either, the detector may see gaps in the stream where a continuous line used to be. Ask the carrier what it does, and test with a recording of the same line.

How do you tune and test VAD for phone lines?

Record real calls, label where speech starts and stops, change one setting at a time, and count both kinds of mistake. Tuning on a clean sample from your laptop will give you numbers that fall apart on a street.

  1. Collect recordings. Gather calls from the places your callers ring from: a balcony, a shop, a bus stop, a living room. Use a mix of fast, slow and soft voices. Get consent and follow your own data rules for the recordings.
  2. Label them. Mark where the caller speaks, to the nearest quarter of a second. A spreadsheet is enough.
  3. Run the detector at the defaults. Write down two numbers: how often noise was called speech, and how often speech was missed.
  4. Change one setting. Move the threshold by 0.05, or the minimum speech by 50 milliseconds. Run again. Keep the table of runs.
  5. Look at the other number. A change that cuts false starts almost always adds misses. Stop when both are tolerable on your worst recordings, not your best.
  6. Retest the whole call. Make a voice test call to the real agent, with the same sounds playing near the phone.

Note that a detector that is too eager and one that is too strict both hurt callers, but in different ways. Eager means the agent stops for scooters. Strict means a soft-spoken caller is not heard and has to say "hello?" twice. Choose the failure your callers will forgive more easily, and test for it on purpose. Our guide on how to test a voice agent covers the wider method.

How Bananaflow fits

Bananaflow is an AI voice agent platform for Indian businesses. The agent's pipeline runs Silero VAD by default to detect when the caller starts and stops speaking, so you do not host the model yourself. Three settings shape turn-taking: barge-in, set for each step of the call; a turn judge, which decides when the caller has finished; and sound sensitivity, low, medium or high, which decides how readily sound on the line counts as the caller talking.

Low ignores quiet or short sounds, which suits noisy lines but can miss a very soft voice. High picks up soft voices, but on a noisy line background sound can hold back the agent's reply. You can test an agent in the browser by voice or text before it takes real calls, and after each call you get a transcript and a recording. Our glossary entry on voice activity detection has the short version, and the languages page lists what the agent speaks.

What should you remember about Silero VAD on phone audio?

  • Silero VAD supports 8000 Hz and 16000 Hz. Phone audio over G.711 is 8000 Hz, so use the 8 kHz setting and 256-sample chunks.
  • The model gives a probability. Your threshold and timing settings turn it into speech or not speech.
  • The defaults are 0.5, 250 ms minimum speech, 100 ms minimum silence and 30 ms padding. Treat them as a start.
  • Traffic gives false starts, soft voices give misses, and a second voice is real speech that is not your caller.
  • Tune on your own recordings, one setting at a time, and count both kinds of mistake.
  • VAD only hears sound. Endpointing and barge-in decide what the sound means.

FAQ

Frequently asked questions

  • What is Silero VAD?

    Silero VAD is an open-source voice activity detector. It takes short chunks of audio and gives a speech probability for each one. Its repository describes it as a pre-trained enterprise-grade voice activity detector, and says it was trained on corpora covering over 6000 languages. It is a model you run yourself, not a hosted service.

  • Does Silero VAD work on 8 kHz phone audio?

    Yes. The repository says Silero VAD supports 8000 Hz and 16000 Hz sampling rates, and 8 kHz is the rate of standard phone audio such as G.711. At 8000 Hz the model takes chunks of 256 samples, which is 32 milliseconds. At 16000 Hz it takes 512 samples, the same 32 milliseconds.

  • What is the default threshold in Silero VAD?

    The default speech threshold is 0.5. Chunks with a speech probability above it count as speech. The repository's own note says it is better to tune the threshold for each dataset, and that 0.5 is pretty good for most. There is also a lower exit threshold, 0.15 below the main one by default, so the detector does not flicker.

  • What do min_speech_duration_ms and min_silence_duration_ms do?

    min_speech_duration_ms (default 250) throws away speech segments shorter than that, which removes clicks and short noises. min_silence_duration_ms (default 100) is how long the detector waits at the end of a segment before it closes it. speech_pad_ms (default 30) adds a little padding on each side so word edges are not clipped.

  • Why does traffic or a TV trigger voice activity detection?

    Because the detector is given a probability, not a fact. A TV serial is human speech, so a good detector may call it speech. Traffic and fans usually give a lower probability, but a loud horn can cross the threshold. Raise the threshold, lengthen the minimum speech, and test on recordings of your own callers' lines.

  • How do you tune VAD for Indian phone calls?

    Record real calls from the places your callers ring from, label where speech starts and stops, and change one setting at a time. Count two kinds of mistakes: noise heard as speech, and speech missed. A setting that fixes one often worsens the other, so keep the numbers for each run and compare.

  • What is the sound sensitivity setting on Bananaflow?

    Sound sensitivity, low, medium or high, decides how readily sound on the line counts as the caller talking. Low ignores quiet or short sounds, which suits noisy lines but can miss a very soft voice. High picks up soft voices, but on a noisy line background sound can hold back the agent's reply. The agent's pipeline runs Silero VAD by default.

Written by

  • Jaseem Abbas, founder of Bananaflow, in glasses and a dark top against a pale background.

    Jaseem Abbas

    Founder, Bananaflow

    Jaseem founded Bananaflow to build voice AI that understands how people in India really speak, including the switch between languages in the middle of a call.

Similar articles

Bring us one process. We'll pilot it on your real calls, with your team as the fallback.

Plan a pilot