HD-Voice for cellular communications - Khariton, Dmitry, Vladimir, Oleg, Irina, Glory!

We turned on HD-Voice on the Beeline network in Moscow. Below are sound recordings before and after, as well as technical details.

The HD-voice mode is based on the AMR-WB technology (Adaptive Multi Rate Wadeband) - adaptive broadband coding with variable speed, this codec is also known as the G.722.22 standard. In theory, a person can hear sounds in the range of 20Hz - 20kHz, but in practice it is enough to transmit sound in a much narrower range. When encoding speech in AMR-WB, the frequency range of 50Hz-7kHz is used, this is absolutely enough for full voice transmission. For example, the AMR codec uses the 300-3 kHz range. The method can be used on 2G and 3G networks.

Simply put, now the human voice will sound more natural, because it will cease to “squeeze” into a rather narrow strip, the need for which was once dictated by the limitations of infrastructure. In addition, with HD coding, external noise has less impact on speech intelligibility.

How did we get to HD-Voice, and how does it work?

A person is able to hear a certain band of sound frequencies, usually it is taken equal to 20Hz-20kHz. It is difficult to transmit the entire spectrum via wireless communication channels without compression, so voice without band cuts is transmitted only in wired systems. For the purpose of speech compression, voice codecs or vocoders are used. Their side effect is the distortion of the transmitted voice by limiting the frequency band of the voice message.

The very first voice codec used on mobile networks was “Full Rate” - FR. Simultaneously with FR, the Half Rate codec was also introduced, its sole purpose is to increase the number of simultaneously served subscribers, and, as its name implies, it occupies half the band of the FR codec in the radio channel. However, the voice quality after transcoding to FR was not very high, and for some countries (for example, Arab, where the speech speed is high and there are a lot of high-frequency sounds), is completely unacceptable. Therefore, the FR codec was redesigned and the Enhanced Full Rate appeared, providing significantly better voice quality with less bitrate being created.

With the growth of mobile network penetration, requirements for the quality of voice services began to grow, and there was a problem with the deterioration of voice quality, with a low level of radio signal from the base station. To solve this problem, a new codec was developed - “Adaptive Multi Rate”, which used the same frequency range for analysis and compression, but the algorithm was implemented in such a way that the bitrate of the codec changed dynamically, depending on the quality of the received signal. Thanks to this, it was possible, with good radio conditions, to provide excellent voice quality, and with bad conditions, to keep a conversation even where FR / EFR codecs could not work at all.

All 3G networks work only using the AMR codec or its subsequent implementations, for example AMR-WB. But since there are still a large number of subscribers using phones that only support EFR / HR codecs, the 2G network continues to work with all codecs (FR / EFR / HR / AMR FR / AMR HR) immediately, providing voice communications to any phone.

Currently, voice quality requirements in mobile networks continue to grow, and it is time to use AMR-WB technology. All codecs described above use a frequency range up to 200Hz-3.4kHz for compression, while AMR-WB uses a voice message bandwidth of 50Hz-7kHz for encoding., I.e. AMR-WB allows you to transmit twice the large frequency band, which gives an increase in the quality and saturation of high and low frequencies of the transmitted voice.

Some codec specifications are listed in the table:

Codec	Standard	Year of creation	Range of compressible frequencies	Created Bitrate
Full Rate - FR	GSM 06.10	1990	200-3400 Hz	13 kbit / s
Half Rate - HR	GSM 06.20	1990	200-3400 Hz	5.6 kbit / s
Enhanced Full Rate - EFR	GSM 06.60	1995	200-3400 Hz	12.2 kbit / s
Adaptive Multi Rate - AMR	3GPP TS 26.071	1999	200-3400 Hz	4.75 - 12.20 kbit / s
Adaptive Multi Rate - WideBand - AMR-WB	3GPP TS 26.190	2001	50-7000 Hz	6.60 - 23.85 kbit / s
Adaptive Multi Rate-WideBand + - AMR-WB +	3GPP TS 26.290	2004	50-7000 Hz	6 - 36 kbit / s (mono) 7 - 48 kbit / s (stereo)

Hint for curious programmers: a sample implementation of the AMR codec in C is available in the 3GPP TS 26.073 standard (you can easily find, download and try to implement it).

In the diagram below, you can see a comparison of voice quality, according to the MOS (Mean Opinion Score) scale, depending on the codec used. The MOS scale is a subjective assessment of voice quality from 0 to 5, where 0 is the complete absence of audibility, and 5 is live speech. G.711 codec - used in landline telephony, that is, this is the quality that you should hear from your landline phone (unless, of course, it has been transferred to the IP channel, but this is a topic for a completely different conversation).



Update: The comments rightly pointed out that the average MOS score for the G.711 codec is about 4.1, this picture slightly distorts reality.
As you can see from the diagram (and taking into account the information from Update), the AMR-WB codec provides, with good radio conditions, voice quality similar to that which we are used to when using a wired home telephone.

For example, two fragments of a musical composition, which were compressed by AMR and AMR-WB codecs, you can compare the sound quality yourself and visually evaluate the difference in the sound picture created by different codecs.

AMR encoded upper track

AMR-encoded lower track

Tests

We are not the first to launch this technology in Russia, so there have already been some developments. Our first network segment with HD-Voice was built within the Moscow Ring Road, using Ericsson equipment. Activating AMR-WB took a long time: at first it was tested on one controller (with numerous drive tests for measuring speech quality). And even after everything was tested on one controller, the subsequent activation took place in stages, controller after controller. The entire activation process took almost six months.

Here are the reference samples from the final tests , which were evaluated in hardware using accurate measurements at different times.

Does it already work?

Yes, now the technology is implemented in a 3G network for Moscow. In order to use the mode, you do not need to make additional settings, it activates itself if both subscribers are on the 3G network and both of their phones support the codec.



The diagram shows the conditions that are necessary for the operation of the AMR-WB. Now, as mentioned above, the codec is launched on a 3G network, the next step is to ensure the codec works between a 2G and 3G network, then between networks and between operators.

What devices are supported?

Recognition problems

Human speech is such that the sounds “C” and “F” are poorly distinguished in the radio and on a narrow band of the audio spectrum. The pairs “T” and “P”, “M” and “H” also suffer. Hence the mnemonic system with the names: "I dictate, Olga’s number two, zero, zero, three, as I understand it, the reception?" Things grow out of here, like assigning the most critical parts of the callsigns with “p” - for example, a sniper, who in no case should confuse his callsign, most likely will be “thirty-third”, because on the air this is very distinguishable in narrow stripes.

Also with HD-Voice voice recognition is growing. This is important for personal communication, and for the growing popularity of additional authentication by voice, for example, in banks.