Telco (voice) channel

" sample rate The sample rate in samples per second (`Hertz' or `Hz'). Digi‐ tal telephony traditionally uses a sample rate of 8000 Hz (8 kHz), though these days, 16 and even 32 kHz are becoming more common. Audio Compact Discs use 44100 Hz (44.1 kHz). Digital Au‐ dio Tape and many computer systems use 48 kHz. Professional au‐ dio systems often use 96 kHz.

The term `bit-rate' is a measure of the amount of storage occupied by an encoded audio signal over a unit of time. It can depend on all of the above and is typically denoted as a number of kilo-bits per second (kbps). An A-law telephony signal has a bit-rate of 64 kbps. MP3-en‐ coded stereo music typically has a bit-rate of 128-196 kbps. FLAC-en‐ coded stereo music typically has a bit-rate of 550-760 kbps.

Accuracy Many file formats that compress audio discard some of the audio signal information whilst doing so. Converting to such a format and then con‐ verting back again will not produce an exact copy of the original au‐ dio. This is the case for many formats used in telephony (e.g. A-law, GSM) where low signal bandwidth is more important than high audio fi‐ delity, and for many formats used in portable music players (e.g. MP3, Vorbis) where adequate fidelity can be retained even with the large compression ratios that are needed to make portable players practical.

a-law International telephony standard for logarithmic encoding to 8 bits per sample. It has a precision equivalent to roughly 13-bit PCM and is sometimes encoded with reversed bit-ordering (see the -X option).

u-law, mu-law North American telephony standard for logarithmic encod‐ ing to 8 bits per sample. A.k.a. μ-law. It has a preci‐ sion equivalent to roughly 14-bit PCM and is sometimes encoded with reversed bit-ordering (see the -X option).

gsm-full-rate GSM is currently used for the vast majority of the world's digital wireless telephone calls. It utilises several audio formats with different bit-rates and asso‐ ciated speech quality. SoX has support for GSM's origi‐ nal 13kbps `Full Rate' audio format. It is usually CPU- intensive to work with GSM audio.

LICENSE Copyright 1998-2013 Chris Bagwell and SoX Contributors. Copyright 1991 Lance Norskog and Sundry Contributors.

AUTHORS Chris Bagwell ( snipped-for-privacy@users.sourceforge.net). Other authors and con‐ tributors are listed in the ChangeLog file that is distributed with the source code." says info sox

Telephone-company employees must sign documentation promising that they will not enact unfaithful propagations of persons' voices, but completely faithful telephones do not exist. (S.

Back in those days - well around year 2000 at least - a rather different model actually applied. This was dc to 4kHz. I tested phone calls between the UK and USA and Canada and also across Europe and was usually able to pass dc across the phone network. This did require using ISDN equipment at each end. Even that far back most of the global phone network was digital and it was only the codecs at each end that defined the frequency response. Many codecs had a high-pass corner at around

But, one can't be sure of how point A and point B are interconnected, even if you know the type of "connection" at one end.

E.g., I can call a neighbor using an old-fashioned land line. Or, call the same neighbor from a cell phone. Or, call the

That's just not practical. If I called you, today, how could

Now, instead of "I" and "you" imagine two indiscriminate individuals with no technical expertise and no *interest* in any of that... Does party A "sound different" (than expectations) because of a head cold? Or, some characteristic of the particular connection?

It's not just that they aren't "faithful" but, rather, that they aren't *repeatable*.

You may be able to infer a rough idea of the channel frequency response by taking an FFT of a chunk digitised at 44kHz. But I'd be more inclined to low pass filter everything above 4kHz and use 8b @ 8kHz.

It should be good enough for the intended purpose of speech at that.

I typically am dealing with speech that is conveyed in open air or via a relatively high bandwidth channel. I continually update my models based on my knowledge of the speaker's identity AND THE "unimposing" CHARACTERISTICS OF THE CHANNEL.

But, also have to handle the likely possibility of a telecom channel for the same speaker.

This both distorts their perceived speech AND would bias the model that has been built without those constraints.

If all you are doing is *recognition*, its not a problem. But, for diarization and identification, it lowers their effectiveness. Lack of sidetone and the latency of modern telco already leaves you with a significant challenge!

As an "inverse filter" can't reconstruct characteristics that have been discarded by a lower sampling frequency, I have two choices:

- run the trained models through an appropriate filter that models the telco channel

- build a separate model for *just* the telco channel

But, if that channel can change from call to call... <shrug>

I *might* be able to use PRESUMED knowledge of the party's identity to help characterize the current channel. But, that would defeat identification.

...

A-law does not compress. It only changes the amplitude. Does not change the bandwidth. A-law changes the amplitude encoding, not the audio bandwidth.

Cheers, Carlos. ES🇪🇸, EU🇪🇺;

To use 8 bits, you normally had to use companding digitizers (A-law over here, mu-law in Europe, iirc).

Haven’t seen one of those in a looong time.

Cheers

Phil Hobbs

The Telco minimum standard is 300 to 3400 hz (or 3200 for some equipment manufacturers, depending on who you talk to). LPF is required before digitizing to avoid aliasing.

In some implementations, such as some special circuit voice lines, the least significant bit is reserved for supervisory signalling.

Cell uses a much narrower bandwidth. They do some data compression to save bits, the nature of which i can't tell you, but the saving is dramatic. You can tell the difference in sound quality, tho, if you're calling from a landline to another landline, as opposed to calling a cell phone.

Yes, but that was when PSTN was *the* means of moving voice. How do more modern network protocols address this? Do they deliberately cripple themselves thinking there is a land-line (over copper, SLIC06, etc.) on at least one end of the line (so why go out of your way to provide MORE fidelity)?

That is exactly the issue.

If you speak into a microphone hard-wired to a device that digitizes your speech at a high frequency and resolution and have models of your speech built in that environment, how do you match those models to speech that has made its way to the analyzer/models via channels of dubious characteristics?

POTS relied on the fact that humans can recognize even severely distorted speech and can rely on knowledge of WHO they telephoned (or who reassured themselves at the far end) to identify the called parties.

Note how some vendors now rely on speech as a biometric authenticator. How reliable could *that* be if I placed a call to them from an old exchange through a dubious PBX, etc.?

The PSTN is still the go-to carrier for voice calls. There are VOIP and various ATM vehicles, which would likely use telco routers and switches or telco TDM facilities... if you need wider sound bandwidth. I'm not sure what more modern protocols you mean. I'm not up to date on that.

Most of the information in a human voice signal is well within the sub-1khz range. More so for men than women.

Not sure what you mean. The only part of the "speech model" that matters is the spectral content. Signal/noise matters, but that's a hardware issue.

I don't know, but probably just fine. One would think they are built for telco channel transmission.

In most cases the PBX or other customer facilities does the A/D and D/A conversion at the customer premise, so analog signals traveling down long copper spans are not a source of degradation.

What's the alternative... a call over Instagram? A high-priced special circuit from your carrier?

There are more ways of moving voice signals from speaker A to listener B than in decades past. Each one "colors" the speech in ways that are easily accommodated by human listeners but regarded as distinct by analytical mechanisms.

What I want is repeatability and, ideally, with some form of quantitative description that lets me relate THIS audio stream to an audio stream that I would encounter when sharing a room with the speaker.

But, that's because the human listener is "accommodating". One can distort their voice and still be understood by someone intent on that understanding. Witness how you can understand someone with a severe head cold, despite their voice being quantitatively "different".

Or, someone with laryngitis.

A machine can similarly understand what they are *saying*. But, its harder for a machine to identify and authenticate that speaker. "What does Bob sound like when he's got a head cold?"

Anything that the channel does to the signal falls into the same problem domain.

E.g., I have a bridge that lets my *wired* home phones use my cellular connection. But, the quality of the connection is beneath the quality of the wired phones on a wired landline *or* the call phone that is connected to the bridge.

But "you" sound different when your speech travels to "me" via different channels. I can understand what you are trying to

They rely on other information -- your caller ID (which can be forged), your knowledge of an account number, etc. And, if their (machine's) confidence in the authentication isn't high enough, they let a human operator perform that task before granting you access to "confidential" information (even though it is YOUR information)

When someone you know phones you, you are able to *identify* the caller without checking their CID or waiting for them to introduce themselves. Because you likely have spoken with them over the telephone (channel) previously and knows what their "telephone voice" sounds like.

I'm not asking (or requiring) anything special. Rather, I am trying to learn the characteristics of a channel that has presented itself to me (my device) without any input from me (or the other party!) in that selection. I.e., you can't tell your provider what sort of connection you want for THIS call.

If I phone home and simply state "Meet me out front", my other half will do so, even if annoyed at the terse, imperative nature of my comment. Because she would

I want a machine to be able to act with that same level of reliability without having to resort to supplemental interactions to bolster confidence in its authentication of the caller.

So, if I am out of town, I can instruct the machine to recognize a neighbor's "commands" (to perform a limited set of actions, as my proxy) without having to quiz them on other information (effectively, shared secrets -- even if not truly "secret") to get that level of confidence in their authentication:

"What day did Don depart?" "When did he say he was going to return?" "Where was he going?" "What's your pet's name?"

I.e., a recording of said neighbor (or, an AI generated simulant) could sound correct and yet not respond correctly.

A good summary of the various technologies involved:

"In 2023 Vice News and The Guardian separately demonstrated they could defeat standard financial speaker-authentication systems using AI-generated voices generated from about five minutes of the target's voice samples."

(I don't think you even need THAT much training material)

Note that the banking application inherently constrains the types of "content" that is likely exchanged making it easier to spoof.

Some mobile 'phones produce sounds that are barely recognisable as a voice, let alone any particular voice. This is made even worse by people who stand them on a table and switch them to 'hands-free' with a loud television or radio in the background.

But that's also true of wired (and "wireless") phones. I had a "single chip" phone many decades ago that felt like I was using a can-on-a-string -- a mild breeze would blow it off the table!

The fact that they can get any sort of audio quality (think: "music player") out of a cell phone suggests they are putting SOME effort into that.

(snip for focus)

That would be signal-to-noise ratio (SNR). SNR = St / (St - Sr) if St is the transmitted signal, ie out of the speaker's mouth, and Sr is the received signal, that which goes into your ear.

Simple to say, accurate measurement could be pretty difficult.

You could send a known signal thru the network that would quantify frequency and phase distortion and latency. I don't know what off-the-shelf equipment would do that, but there must be some.

No. It's because that's just the spectral content of a human voice. That one can be objectively measured & quantified.

Not exactly. What the channel does to the signal is noise, which falls under SNR. If Bob has a head cold, that's just a difference in the original signal.

If you are trying to identify the caller via a "voice print", it still adds to the confusion all the same. IIRC a voice print can't be spoofed by another human voice, but i'm pretty sure it can be spoofed by an AI synthesis of Bob's voice. I speculate the voice print would not be greatly affected if Bob has a head cold, since it sounds like a product of the speaker's physiology, like larynx, tongue etc, at least according to my very incomplete understanding.

That would be a pretty serious signal degradation.

(snip)

I see.

To the extent that a phone call is digitized the entire way, except for the first and last mile (or fraction thereof), different call paths aren't going to make any difference.

It's also probable that, if you use the PSTN, you are going to get the same path every time anyway. The PSTN mostly uses TDM transport which is pre-determined, rather than multiple paths like the Internet does.

(snip the rest, as I have nothing useful to add to it)

I think it is more than that. I think the channel "colors" the content.

What prompted the post was a demo I gave the other day. With my ear buds in place, I was successfully authenticated (context independant).

I borrowed one of my guests' cell phone and then placed a call to "myself". And, was NOT authenticated. My voice hadn't changed in the few minutes between those two tests. I didn't choose a different utterance.

What had changed was the channel between my mouth and the recognizer.

When I had initially designed the mechanism, I verified it with a POTS connection. Now, the station set was replaced with a cell phone and the copper line with a packet switch network operating over a wireless connection.

To the machine, my voice was no longer strictly what it should have been.

In theory, KNOWING the identity of the remote party AND recognizing whatever utterance, I could dynammically determine an inverse function that maps what I'm *hearing* from him to what he is likely *saying*. That function would qualify the channel's characteristics.

But, if I want to normalize the channel SO I CAN AUTHENTICATE THE CALLER, I've already corrupted that analysis by using my knowledge of his speech to adjust the channel so it *matches* his speech. That's rife with fallacy.

But only if it is faithfully presented to the analyzer. If it is altered in transit, then all bets are off. (Well, you can claim to make a match by enlarging the confidence interval you're willing to tolerate for "certainty")

Shirley the instrument used doesn't differ from every other instrument that could have been used to convey the signal? I.e., all cell phones, carriers, etc. are 100% interchangeable. The mechanical dimensions of transducers don't attenuate or amplify certain frequencies, etc.

It's just a fluke that the cell phone used in my test (above) happened to corrupt the signal delivered to the network?

Your physiology changes in each of these "abnormal" situations.

When you "get a cold", your larynx swells which drives the fundamental frequency of all of your voiced sounds lower. They can become coated with mucus further deepening the voice. Coughing or clearing phlegm irritates the vocal chords making your voice sound more raspy or "shredded". The nasal passages tend to get blocked or narrowed reducing the prominence of unvoiced sounds (m, ng, etc.) as the resonance is altered by the physical changes to that "chamber". Soft tissues don't vibrate as easily. etc.

Most "impersonators" rely on other mechanisms to reinforce the perception of their voice. A machine can note the difference in spectral content while a human listener is more willing to tolerate deviations.

When I am over-tired, my voice gets very breathy and more monotonic. A human can understand what I am saying -- as can a machine. But, expecting the machine to correctly

An AI (or other synthetic voice) is actually using a model built *from* the speaker's voice so it more accurately reflects the original speaker's voice.

To defeat an AI, you have to posses one or more "secrets" that the AI can't deduce from the situation. Or, hope the AI is sluggish in responding to your prompts (which can be spoken or visual, etc. -- "How many fingers am I holding up?")

There are no standards for the station sets on each end of a comm link. Because humans can recognize seriously degraded speech, fidelity isn't an important criteria.

"ATM is a core protocol used in the synchronous optical networking and synchronous digital hierarchy (SONET/SDH) backbone of the public switched telephone network and in the Integrated Services Digital Network (ISDN) but has largely been superseded in favor of next-generation networks based on IP technology."

IP doesn't ensure that one datagram follows the same path as the preceding or following.