The DSP Speech Processor was designed to support a 4 kHz speech bandwidth to suit the transceiver it was designed with (although it also supports narrower bandwidths). Wider speech bandwidths have been found to improve both speech quality and intelligibility in some noisy conditions (particularly QRN). This goes against the old thinking that a narrow bandwidth is always more readable through noise.
Narrow transmit bandwidths concentrate power in a narrower range, causing the signal to stand out above the noise more. But wider speech bandwidths preserve more speech information, which can enhance intelligibility. This is even true of bass frequencies (< 600 Hz), that are clearly audible through noise and give the listener clues to speech timing and pronunciation.
For this to work, good EQ is needed to keep the power spectrum reasonably flat. Without good EQ, the strong bass frequencies will steal most of the available transmitter power, producing a high average power output with very poor readability. Dynamic pre-emphasis, multiband compression and multiband clipping all help achieve a flatter power spectrum. This can result in a signal that sounds like it has too much high frequency energy, but receiver de-emphasis or a high-cut filter corrects this easily, and improves the perceived S/N ratio.
Increasing the bandwidth from 300 to 3000 Hz to 100 to 3000 Hz is an increase of 200 Hz, or less than 8%. If the power spectrum is kept flat, this reduces the power of the frequencies from 300 to 3000 Hz by only 0.7 dB (worst case). On a receiver with de-emphasis (and adequate bass response), the signal can appear to get louder, and more natural sounding.
SSB and AM systems both have a flat noise spectrum. This is because the transmission bandwidth is a small fraction of the radio frequency used, and the modes themselves respond equally to noise across the passband. SSB systems perform a linear translation of frequencies from audio to RF and from RF back to audio. This means as the receiver IF filter bandwidth is increased, the new noise introduced falls on new frequencies. The total noise power in the passband increases, but there is more passband - so the signals stay above the noise floor by the same amount. This is quite easy to observe using an SDR receiver spectrum display. This means the IF bandwidth of an SSB receiver can be increased, without degrading the readability of signals! The total noise may sound louder, but the signals will be just as audible as they were before, as they are the same height above the noise floor. Increasing receiver IF bandwidth also reduces ringing on noise impulses, which can reduce their effect.
Simple diode AM detectors respond to signals and noise across the passband and will degrade the S/N ratio at all frequencies as the bandwidth is increased. A solution to this is to use synchronous AM detection. This helps improve S/N ratio and reduce distortion and is frequently used for these reasons in the video IF detectors of analog TV receivers (remember those?).
Most speech power falls on specific frequencies, rather than being a perfectly flat spectrum like white noise. This causes the transmit power to concentrate in peaks across the passband. This increases the S/N ratio of the speech by making it rise higher above the noise floor. This allows SSB transmitter bandwidth to be increased with a smaller noise penalty than expected.
Future development in speech processing using
FFT methods may allow improvements by modifying the spectrum of the voice
to optimize the spectral peaks being transmitted for higher intelligibility,
while simultaneously modifying their relative phases to produce a lower
peak-to-average output waveform. This may also allow very high loudness
with less distortion than current methods.
To hear the results of these tests accurately, the playback system must have adequate bass response. Good quality headphones are recommended.
UPDATE: A narrowband receiver simulation has been added, with sharp FFT filtering removing all frequencies outside of the range 250 Hz to 3 kHz.
Some observations after much testing:
Tested 10/08/2014:
The Mode 5 tests were
done using a contoured audio EQ out of the multiband compressor, reducing
bass energy between 125 and 600Hz, and boosting treble above 1kHz.
Normal clipping, Mode 5 + EQ boost @ -5 dB SNR - Utility modes stepped through to change bandwidth
Maximum clipping, Mode 5 + EQ boost @ -6 dB SNR - Utility modes stepped through to change bandwidth
Bass/Nobass test @ -6 dB SNR This test compares a highpass cutoff of 60 Hz and 250 Hz at a poor S/N ratio of -6dB.
Very weak Bass/Nobass test @ -10 dB SNR This test compares a highpass cutoff of 60 Hz and 250 Hz at a very poor S/N ratio of -10dB.
Bass/NoBass test - Clean signal, maximum clipping and EQ
Flat
spectrum (Mode 3 + input EQ), maximum clipping, diminishing S/N ratio.
60 Hz - 3 kHz and 250 Hz - 3 kHz, in a loop. S/N ratios: Max, -6 dB, -8
dB, -10 dB, -12 dB, -14 dB.
The results are close.
The 250 Hz highpass may have slightly better readability, particularly
on words "3, 4, 5".
Note: Since these tests were done, the mode numbering has been rearranged. Mode 4 now has a contoured output: -10dB@300Hz, +3dB@1.5kHz and Mode 5 has a flat output.
The tests below have
the new mode numbering:
New:
Mode 4 + Maximum Clipping at -6dB SNR with post EQ and lowpass filtering
at 3kHz Wide:60Hz-3kHz and Narrow:250Hz-3kHz
Which is louder? Which
is more readable? There isn't much difference, except in tonal quality.
New:
Mode 5 + Maximum Clipping at -6dB SNR with post EQ and lowpass filtering
at 3kHz Wide:60Hz-3kHz and Narrow:250Hz-3kHz
Which is louder? Which
is more readable? There isn't much difference, except in tonal quality.
References:
"Spectral redundancy: Intelligibility of sentences heard through narrow spectral slits" 1995
"The
Effect of Bandwidth on Speech Intelligibility" 2006
Hilbert clipping sound quality variations due to changes in EQ.
A flat power spectrum is desirable, for keeping readability high and distortion low. The average power spectrum won't be ruler-flat, due to higher voice frequencies having lower average power, and pronunciation causing power spectrum changes.
With moderate clipping levels, small variations from flat can improve quality. Less clipping in the range 300 to 600 Hz sounds better, as distortion here sounds particularly "muddy". However, this also reduces loudness. Slightly more clipping in the range 1 to 2 kHz produces a stronger clarity on pronunciation, and distortion here seems far less obvious.
This contoured response is produced in Mode 4.
The modified Hilbert clipping routine seems to tolerate higher level at the voice fundamental frequency (100 Hz here), possibly because the distortion products are masked by the existing voice harmonics. Using a good amount of phase rotation can reduce clipper distortion and raise average level. This is due to the phase rotation causing an improvement in speech waveform symmetry, and temporal skewing of the spectrum causing the different voice frequencies to overlap less, and hence clip at different times. If heavy clipping is used for DX purposes, a totally flat power spectrum (adjusted using white noise, with all multiband compressors in gain reduction) produces louder and better sounding audio.
This flat output is produced in Mode 5.
Since Mode 5 is louder and has good quality over typical communications channels, it is now the processor's default mode.
