Math Bridge: Quantization and Quiet Speech

← Back to Voice and Audio Compression for IoT
Math BridgeOne learning thread

Why quiet speech needs companding

Separate sample rate from bit depth, then reproduce the chapter’s 4 kHz ceiling, 49.9 dB baseline, and 9.9 dB quiet-talker result.

Phoebe guides this bridge
One targetExplain why fixed linear steps punish quiet speech.
Chapter case8 kHz, 8-bit, 64 kbps narrowband voice.
What it buys youSee what companding changes without adding bits.

See the relationship before changing it

The chapter figure below shows: Codec choice moves along a payload/fidelity axis, but every option still needs packet, latency, loss, CPU, energy, and listening validation.

The retained audit below checks several chapter fixtures. This model keeps those stated values fixed and changes only voice quantizer resolution, so the numeric fixture does not switch without explanation.

Codec family map comparing PCM, companded audio, hybrid modes, and low-rate speech-model codecs against fidelity and payload size.
Codec choice moves along a payload/fidelity axis, but every option still needs packet, latency, loss, CPU, energy, and listening validation.

Derive the baseline in four named moves

  1. 1

    Name the input. The chapter baseline is 8 bits.

  2. 2

    Name the relationship. quiet-speech step = 2,000 mV / 2^bits

  3. 3

    Substitute with units. 2,000 / 2^8 = 7.813 mV

  4. 4

    Read the result. Keep the unit beside the value. Use it only inside the technical boundary on this page.

Predict, then change voice quantizer resolution

Try Predict the direction of quiet-speech step = 2,000 mV / 2^bits. Test another voice quantizer resolution, then compare ideal amplitude step.

8 bits
Chapter baseline
Ideal amplitude step

Observe Extra bits reduce ideal quantization steps before microphone noise is counted. Reset voice quantizer resolution to 8 and compare ideal amplitude step.

Explain More quantizer bits divide the same voice-signal voltage span into exponentially more levels, so each added bit halves the ideal amplitude step.

Check yourself

What should you do before trusting a moved-control result?
Answer: Predict its direction, apply the shown relationship, keep the units, and reset to the worked baseline.
What does this small model leave out?
Answer: Only voice quantizer resolution moves here. Field effects named in the technical boundary stay fixed.

Honesty note. This teaching model changes only voice quantizer resolution. It is not a field prediction.

1. Separate time from amplitude

Sample rate says how often voltage is read. Bit depth says how many amplitude codes can store each reading. An 8 kHz, 8-bit stream takes 8000 readings each second and gives each reading one of 256 levels.

Phoebe: More readings do not make the amplitude ruler finer; more bits do not make the clock faster.

2. Find the 4 kHz content ceiling

fmax = fs/2
1

Substitute the chapter ratefmax = 8000/2 Hz.

2

Dividefmax = 4000 Hz = 4 kHz.

This is the ideal Nyquist ceiling, not a promise of flat microphone or codec response to 4 kHz.

3. Build the linear quantization floor

Uniform rounding error from step size q lies between −q/2 and +q/2. Its RMS value is q/√12. For a full-scale sine, that produces the familiar ideal estimate:

SNRdB = 6.02N + 1.76
1

Substitute 8 bits6.02(8)+1.76.

2

Multiply48.16+1.76.

3

Add49.92 dB ≈ 49.9 dB.

4. Compare the chapter’s 16-bit figure

1

Substitute 16 bits6.02(16)+1.76.

2

Multiply and add96.32+1.76 = 98.08 dB ≈ 98.1 dB.

3

Compare98.1−49.9 = 48.2 dB more ideal full-scale SNR, while the chapter’s bit rate rises from 64 kbps to 128 kbps.

5. Move the talker 40 dB below full scale

The linear quantizer’s absolute step stays fixed. A signal L dB below full scale loses L dB of SNR.

SNRdB(L) = 6.02N + 1.76 − L
1

Use the 8-bit baselineSNR(40) = 49.9 − 40.

2

SubtractSNR(40) = 9.9 dB.

3

Change the ruler, not the wire widthμ-law and A-law compress large amplitudes before 8-bit quantization and expand them after decoding, creating finer effective steps near quiet speech while still carrying 8 bits at 64 kbps.

6. Check yourself

1. What content ceiling follows from 8 kHz sampling?

4000 Hz, or 4 kHz.

2. What ideal full-scale SNR follows from 8 bits?

6.02(8)+1.76 = 49.9 dB.

3. Why doesn’t companding need a ninth wire bit?

It redistributes effective quantization resolution toward quiet amplitudes before the same 8-bit code is sent.

7. Honesty boundary

The amplitude-step calculation fixes a 2000 mV span, and the ideal signal-to-noise estimate assumes a uniform quantiser, a full-scale sine, and uncorrelated rounding noise. Microphones, gain control, codecs, packet loss, and perception add limits, while a voice 40 dB below full scale is only an illustrative case.

6.02N+1.76 result assumes an ideal uniform quantizer
Ideal uniform-quantiser signal-to-noise formula
a full-scale sine
Assumed input waveform at full amplitude
uncorrelated rounding noise
Assumed uncorrelated quantisation error
Real microphones
Unmodelled microphone response and noise
gain control
Unmodelled automatic amplification effects
codecs
Unmodelled speech encoding effects
packet loss
Unmodelled missing audio packets
hearing perception set additional limits
Perceived speech quality needs separate evaluation
“40 dB below full scale” is a realistic illustrative case named by the chapter
Illustrative quiet-speech level relative to full scale
not a universal quiet-talker level
Illustrative level does not represent every quiet talker

Go deeper in From Samples to a Speech Model for predictive codecs and perceptual trade-offs.