- coefficients collectively make up the MFC.
- derived from a type of cepstral Representation of the audio clip ( Nonlinear “spectrum of a spectrum” )
- commonly used for speech recognition feature systems, for example systems that an recognize a spoken number over a telephone.
we should test if MFCCs are a valid algorithm for HealthTRAC project, and evals on the Temporal Dynamics
smithSpeechAcousticMarkers2025 (Smith et al., 2025)
10-29-2026 Nemotron 3.5 Transcribe
This new model actaully uses a log mel spectrogram for encoding waveforms for the transformer to process!
- Rich transcription: 10 timed event tags (laughter, breath, throat/cough, effort/strain, crying, mouth noise, pause, shout, gesture noise, sung vocals) and 11 speaking-style spans (flat, vocal texture/delivery, vocal intensity/pitch, sadness/regret, calm/reflective, uncertainty/surprise, fear/urgency, positive affect, anger/contempt, social intent, authority/assertion). The word-timing model’s vocabulary includes these tags, so they should get times. NVIDIA released this with no quantitative evaluation, trained on 2.3k annotated hours.
- Other prompted tasks: the transcript with disfluencies removed, transcribing only the speaker nearest the microphone, keyword-boosted ASR, language ID, and question answering about speech and sound.
- Measured accuracy: gender 88%, intent 90%, emotion 60%, age 30% (unusable).
- Fine-tuning is where the clinical value is. The report fine-tuned 17 tasks at 0.5–2.3 GPU-hours each:
Bibliography
- Smith, K. M., Williamson, J. R., & Quatieri, T. F. (2025). Speech Acoustic Markers Can Detect Mild Cognitive Impairment in Parkinson’s Disease. IEEE Journal of Selected Topics in Signal Processing, 19(5), 731–740. https://doi.org/10.1109/jstsp.2025.3620716