28 Voice Compression: Codec and Link Constraints

28.1 Start With the Decision

A door intercom can sound clear on the bench but break into gaps on its radio link. Format, loss, and delay must fit the same path.

28.2 Route Overview

This is part 1 of 2. Continue with Voice Compression: Samples, Companding, and Speech Models.

28.3 Part Objectives

• Test why voice gets compressed with a concrete scenario and pass criteria.
• Validate apply it: match the codec to the link with a concrete scenario and pass criteria.

28.4 Chapter Roadmap

• In 60 Seconds
• Start With the Story
• Why Voice Gets Compressed
• Apply It: Match the Codec to the Link

28.5 In 60 Seconds

Fit One Clear Message Through the Real Link

Picture a door intercom that sounds clear on the bench but breaks into gaps over its radio link. The audio format, packet overhead, loss, and delay must fit the same path.

Analog-to-digital converter means a circuit that turns microphone voltage into numbers; it is often shortened to ADC. Latency means the delay between speech and useful playback. Measure both the encoded stream and the listening result.

Send one known phrase at two quality settings, then add loss and a slow link. Keep sample settings, encoded size, packet overhead, delay, losses, energy, and listening notes.

This runway does not prove that one codec fits every voice task. The deeper sections explain sampling, quantization, compression, packet framing, link budgets, loss handling, and listening validation.

Voice compression turns a raw microphone sample stream into a stream small enough for constrained IoT links. Review it as a complete link-budget, packet-overhead, latency, loss, CPU, energy, and listening-validation decision.

28.6 Start With the Story

You will compare audio codecs against a PCM baseline and include packet costs when judging whether voice fits a link. Start with the speech task, capture quality, and available transmission budget.

Follow one voice message across four beats to see how codec, overhead, and delay must fit the same constrained link.

The mathematical gist. At the chapter’s $8$ kHz rate, Nyquist gives a $4$ kHz content ceiling. Ideal linear PCM gives $6.02(8)+1.76=49.9$ dB full-scale SNR, but a talker $40$ dB below full scale keeps only $9.9$ dB. Companding reshapes the amplitude ruler so quiet speech receives finer effective steps while the wire still carries $8$ bits at $64$ kbps.

Math Bridge · guided foundationsWhy does quiet speech need companding?Phoebe separates sample rate from bit depth and works the chapter’s 9.9 dB quiet-talker case.Learn the maths

28.7 Why Voice Gets Compressed

Raw digital audio is bandwidth-hungry. Capturing speech as plain samples produces a steady stream of bits that can overwhelm the narrow, low-power links many IoT devices use. Voice compression shrinks that stream so speech fits the link.

The path is familiar from sampling: a microphone signal is sampled and quantized into numbers, and then compressed before it is packetized and sent. Compression sits after the same sampling and quantization steps covered in the ADC chapter.

Lossy compression is the norm for voice. Unlike a file you must reconstruct exactly, speech only needs to stay intelligible and natural enough, so voice codecs discard detail the ear barely notices in exchange for a much smaller stream. Think of shorthand note-taking: you do not capture every word, only enough to reconstruct the meaning.

For example, a battery-powered site intercom may capture narrowband speech from a MEMS microphone, send it over Wi-Fi or LTE-M backhaul, and need a supervisor to understand a short spoken request with little delay. Raw 8 kHz, 16-bit, mono PCM is 128 kbps before packet headers, and classic 8-bit narrowband PCM is still 64 kbps. A speech codec such as AMR-NB, Opus configured for speech, or Codec2 can reduce the carried voice stream by modelling speech and tolerating small perceptual losses. The choice is not only the advertised codec bit rate; the device must also budget packet headers, jitter buffer delay, dropped packets, CPU energy, and whether the audio is live talk, a push-to-talk clip, or a stored alert.

Before Why Voice Gets Compressed, inspect the figure Figure 28.1. Compare Speech with Capture; their difference reveals Review voice as a complete link: speech task, capture chain, PCM baseline, codec mode, packet budget, and real-speech validation. This gives Why Voice Gets Compressed evidence to revisit.

Follow the speech task across Figure 28.1 through microphone capture, the PCM baseline, codec framing, and packet budgeting. The validation strip below checks real speech, delay, loss, and noise across the complete link.

The One-Minute Voice-Coding View

Beginner Examples

• Plain uncompressed speech samples can exceed what a narrow link comfortably carries, which is why phones compress.
• A one-way recorded alert tolerates more delay than a live two-way conversation.
• Working on bench Wi-Fi does not prove a codec works over a constrained field link.

Voice Compression Knowledge Check

If this gives you the trade-off, you can stop here. Continue to Practitioner when you need to choose a codec and packet size for a real link.

28.8 Apply It: Match the Codec to the Link

The practical job is to pick a codec and bit rate the link can sustain, size the packets to balance overhead against latency and loss, and confirm the result by listening over the real link.

A practical review starts with the worst hour of the deployment, not the best bench measurement. A Wi-Fi door station may have enough throughput but suffer roaming gaps and jitter when the access point is busy. An LTE-M tracker may have enough coverage for telemetry but limited uplink budget for live speech. A phone-tethered wearable may depend on BLE connection interval and app foreground behavior. For each case, test the same sentence, background noise, and packet-loss pattern with the candidate codec and packet size, then log perceived intelligibility, round-trip delay, and gaps.

Walkthrough: From Link Budget to Validated Audio

1. Measure the link budget. Find the reliable, sustained bit rate and the typical packet loss of the real link, not the peak or the lab figure.
2. Set the latency budget. Decide the maximum acceptable end-to-end delay; interactive talk needs far less than one-way playback.
3. Choose the codec family. Waveform coders preserve general audio at higher bit rates; speech-model coders reach much lower bit rates for voice specifically. Match the family to the content and the bit rate the link allows.
4. Size the packets. Each packet carries header overhead, so very small audio packets waste bandwidth on headers, while large packets add delay and lose more audio per dropped packet.
5. Plan for loss. Expect dropped packets on real links; prefer codecs and packet sizes that degrade gracefully, and consider packet-loss concealment.
6. Validate with real speech. Listen to representative voices and conditions, because tone tests do not capture intelligibility or conversational delay.

Read Figure 28.2 from waveform-based PCM toward the low-rate speech models. The axis shows the trade-off between waveform fidelity and payload size; the lower note lists costs beyond codec bit rate.

Compare PCM with companded audio in Figure 28.2: G.711 changes quiet-signal steps while retaining the same 64 kbps rate. Continue right to the speech-model options, where smaller payloads require checking whether the resulting speech meets the task.

The Packet-Sizing Trade-Off

There is no single best packet size, only a balance for your link's bit rate, loss rate, and delay budget.

The The Packet-Sizing Trade-Off argument uses Figure 28.3 to compare Voice Release Record. Look next for PCM Baseline before accepting A release record proves the full path, not just the advertised codec bit rate as a design claim.

Trace Figure 28.3 through Voice Release Record, PCM Baseline, and Packet Budget. At the first stop, the diagram uses Voice Release Record to hold review evidence; at the second it highlights PCM Baseline; at the third it uses Packet Budget to show the next hand-off. Those hand-offs make A release record proves the full path, not just the advertised codec bit rate actionable within The Packet-Sizing Trade-Off.

Incremental Practice

Packet Sizing Knowledge Check

If your job is to choose a codec and packet size, you can stop here. Continue to Under the Hood for the baseline math and the coder families.

28.9 Continue to the Next Part

Carry this evidence into Voice Compression: Samples, Companding, and Speech Models, which begins with Under the Hood: From Samples to a Speech Model.