Keep in mind, cellular carrier stations (cell towers) don't just become congested; they're
oversubscribed, pretty much all the time. Each level of cell service has essentially an SLA for minimum voice quality per line—which can be modelled as each customer having a fixed bandwidth allocation. So, at a given service tier, you can sacrifice latency (which doesn't work for real-time protocols) or drop packets (which is what ends up happening), but you
can't just lower the available bandwidth of the channel. The [hardware!] voice codecs are built to assume the bandwidth when they're on a given cellular network standard—they're essentially emulating circuit-switching!
All that being said, that's why cell networks, until 4G, used actual circuits for voice, rather than having it be packet-based; and that's why 4G+ networks use IMS (a layer 1+2+3 protocol suite that all middle boxes are aware of for QoS purposes) rather than just relying on layer-3 https://en.wikipedia.org/wiki/Differentiated_services QoS fields (which are handled by full layer-3 switches within a given LAN, but get dropped the moment the packet crosses onto the public Internet and are just ignored by layer-2 switches.)
(And even on large LANs—like your cable provider's ATM network—DiffServ tags aren't relied upon, because QoS is just so much easier when layers 1 and 2 are aware of it. If you get a IP phone line from your ISP, it'll be using https://en.wikipedia.org/wiki/Multiprotocol_Label_Switching tags instead, to give its packets QoS priority across even the dumbest middleboxes.)