Why architecture matters here
The architecture matters because the cost it removes is both large and continuous, and the artifact it prevents is both subtle and call-ending. On the cost side, transmitting full speech frames through silence means paying peak bitrate for the quiet majority of a conversation. For one call that is wasteful; for a mobile carrier multiplexing thousands of voice channels over shared spectrum, or a conferencing platform fanning one speaker to hundreds of listeners, the silence traffic is a first-order capacity and cost problem. DTX turns that quiet majority from full-rate traffic into almost nothing, which is why it is standard in every serious voice codec rather than an optimization.
On the quality side, the reason you cannot just stop sending is that human listeners actively use background noise as a liveness signal. A real phone line, a real room, has an ambient floor, and when it vanishes completely the listener's instinct is that the call dropped — they say 'hello? are you there?' into what is actually a perfectly healthy connection. Comfort noise exists because the absence of noise is itself an artifact, and a jarring one. The architecture's job is to make silence sound like silence-with-someone-there, not like a disconnection, and that requires synthesizing rather than transmitting the hush.
It matters, too, because the boundaries between speech and silence are where the whole scheme succeeds or fails perceptually. Speech does not start and stop cleanly; words trail off into low-energy tails, and utterances begin with soft onsets. A DTX system that cuts transmission the instant energy drops will clip the ends of words and swallow soft consonants, and a receiver that switches to comfort noise a beat too late or resumes speech a beat too early produces audible discontinuities at every turn. The value of DTX is entirely conditional on these transitions being handled with hangover and smoothing so the listener never hears the seam — a raw energy gate makes the savings but destroys the experience.
Finally, the architecture matters because it interacts with everything downstream of it. During DTX silence, few or no packets flow, which means the receiver's jitter buffer has nothing to clock against, RTP sequence numbers jump across the gap, and any component that assumes a steady packet cadence must be told that a silence gap is expected rather than a loss. Comfort noise is not just a perceptual nicety; it is the thing that keeps the receiver's timing and loss-detection machinery sane when the sender deliberately goes quiet. Designing DTX means designing how the rest of the pipeline behaves when packets legitimately stop.
The architecture: every piece explained
Top row: the decision and the fork. Incoming mic frames — typically twenty-millisecond blocks of audio — are fed to a voice activity detector (VAD) whose single job is to classify each frame as speech or silence. This is the hinge of the whole design: a good VAD makes DTX both efficient and transparent, a bad one either wastes bandwidth (too eager to call things speech) or clips words (too eager to call things silence). When the VAD says speech, the frame is encoded normally into full codec packets and sent as usual. When it says silence, the encoder engages DTX and stops sending full frames — the transmission goes discontinuous.
Middle-left: the silence-insertion descriptor. Going fully silent on the wire would leave the receiver blind to the background, so during a silence period the sender still emits an occasional SID packet — a silence-insertion descriptor. A SID is tiny: it carries the parameters the receiver needs to reproduce the room's ambient noise, primarily its energy level and a coarse spectral shape (often expressed as filter coefficients). It is sent at the start of a silence period and then refreshed only occasionally — every few hundred milliseconds — just often enough to follow a background that drifts, like an air conditioner cycling or traffic swelling. Between SIDs, the network carries few or no packets at all, which is exactly the savings.
Right column: detection and synthesis at the receiver. The receiver detects the gap and, critically, interprets 'no speech frames arriving' as expected DTX silence rather than a dead line or packet loss — a distinction it can only make because DTX is a known, negotiated mode. It then runs comfort noise generation (CNG): a small synthesizer produces shaped noise matching the level and spectrum described by the most recent SID, filling the silence with a hush that sounds like the speaker's actual background. The listener hears a live, quiet room; the network carried almost nothing to produce it.
Bottom rows: the transitions. Because word tails are low-energy, a raw VAD would cut them off, so the sender applies hangover — it keeps transmitting full frames for a short window after the VAD first calls silence, and smooths the handover so the tail of a word is never clipped. On the other side, when speech resumes, the receiver switches cleanly from comfort noise back to decoded full frames, ideally with a brief crossfade so the transition from synthesized hush to real audio is seamless. The ops strip names the tuning that makes DTX transparent: VAD thresholds set for the environment, a SID refresh cadence fast enough to track the room but slow enough to save bandwidth, careful level matching so the comfort noise is neither louder nor quieter than the real background, and explicit handling of the clock and sequence-number gap that silence introduces.
End-to-end flow
Trace a few seconds of a phone call through a DTX/CNG pipeline as the speaker finishes a sentence, falls silent, and starts again.
Active speech: the speaker is talking. Every twenty-millisecond frame is classified as speech by the VAD, encoded at the codec's normal rate, and sent as a full packet. The receiver decodes and plays them; the channel is running at full bitrate, as it should during actual speech. Nothing about DTX is visible yet.
The tail of a word: the speaker finishes 'okay' and the energy drops as the vowel trails off. The VAD's raw decision flips to silence, but hangover keeps the encoder sending full frames for a short window so the soft tail of the word is transmitted intact rather than clipped. Only after the hangover window elapses with continued silence does the encoder actually engage DTX. This is the difference between a natural word ending and a chopped one.
Entering silence: with DTX engaged, the encoder measures the background — the quiet hum of the speaker's room — and emits a single SID packet describing its level and spectral shape, then stops sending. The network now carries essentially nothing for this channel. The receiver sees the speech frames stop, recognizes the pattern as DTX silence rather than a fault, reads the SID, and begins generating comfort noise: a shaped hiss at the level the SID specified. The listener hears the speaker's room, quiet but present.
Silence continues: the speaker stays quiet for two seconds. The sender transmits nothing except a refreshed SID every few hundred milliseconds, so that if the background changes — a door opens, the AC kicks on — the comfort noise can track it. The receiver keeps synthesizing the hush from the latest SID. Across two seconds of silence the channel carried a few small SID packets instead of a hundred full frames, and the listener never sensed a gap.
Speech resumes: the speaker starts a new sentence. The VAD immediately classifies the onset as speech, the encoder resumes full-frame transmission, and the receiver switches from comfort noise back to decoded audio — ideally with a brief crossfade so the handover from synthesized hush to real voice is inaudible. The one risk at this boundary is a clipped soft onset if the VAD is slow to fire; a well-tuned detector triggers on the leading edge so the first phoneme is captured. From the listener's perspective the conversation was continuous the whole time; from the network's perspective, most of it cost almost nothing to carry.