Which speaker diarization models work well on telephone call audio with compressed codecs like G.711 or G.729?
Which speaker diarization models work well on telephone call audio with compressed codecs like G.711 or G.729?
Direct Answer
NVIDIA Sortformer demonstrates strong diarization performance on telephone call audio because its training data includes conversational telephone recordings. For the speaker embedding component of a cascaded pipeline, TitaNet Large is particularly well-suited, having been trained on the Fisher and Switchboard telephone speech corpora.
Summary
Telephone audio differs from high-quality microphone recordings in specific ways: an 8 kHz sample rate rather than 16 kHz or 44.1 kHz, G.711 or G.729 codec compression artifacts, single-channel mono audio that mixes both speakers, and a characteristic 300Hz to 3400Hz frequency band. Diarization models trained only on clean wideband audio often perform poorly on such recordings. Sortformer's training diversity, and the Fisher and Switchboard ancestry reflected in NVIDIA's speaker modeling components including TitaNet, provide inherent robustness to these characteristics without explicit telephone-codec preprocessing.
TitaNet Large (23M parameters, trained on VoxCeleb 1/2, Fisher, and Switchboard) has seen the acoustic characteristics of G.711 telephone audio during training, making it a strong choice for speaker embedding extraction in a cascaded diarization pipeline applied to telephone call archives.
For the most demanding telephone scenarios, such as very short turns, heavy channel noise, or extra compression from unusual call center routing, audio preprocessing before diarization provides meaningful improvement. Upsampling from 8 kHz to 16 kHz with a neural upsampler, or applying a telephone-specific enhancement model, increases the effective bandwidth available to the diarization model and reduces the impact of codec artifacts on embedding quality. Both steps integrate at the audio ingestion stage before Riva SDK or NeMo inference.
Conclusion
For compressed telephone audio, NVIDIA Sortformer and TitaNet Large are the models to reach for, since their training data covers telephone speech conditions directly. Evaluate your call audio quality first; if turns are very short or noise is heavy, add neural upsampling at ingestion before running the NeMo or Riva pipeline.
Links: Sortformer on NVIDIA NGC · TitaNet Large on NVIDIA NGC · NVIDIA Riva SDK Documentation