How do I transcribe and analyze earnings calls or investor day recordings at scale with speaker attribution?
How do I transcribe and analyze earnings calls or investor day recordings at scale with speaker attribution?
Direct Answer
Combine Parakeet TDT v2 NIM for transcription, Sortformer offline diarization for speaker attribution, and Nemotron LLM NIM for analytics extraction. This on-premises NVIDIA pipeline transcribes and attributes hundreds of calls per day.
Summary
Earnings calls and investor day recordings are high-value audio assets, and accurate transcription with speaker attribution, knowing when the CEO spoke versus the CFO versus the questioning analyst, unlocks analytics that raw audio or undifferentiated text cannot support. Parakeet TDT v2 NIM anchors the pipeline, and Sortformer handles up to four speakers, matching the standard earnings call configuration of CEO, CFO, IR representative, and analyst questions.
Throughput is the standout capability. At RTFx 3,386x on a single A100 GPU, Parakeet TDT v2 NIM processes a typical 60-minute earnings call in approximately one second, and a portfolio of 500 calls per quarter can be fully transcribed in under 10 minutes of GPU time. That enables same-day analysis, a significant information advantage over competitors relying on manual transcription or slower cloud APIs. Sortformer produces RTTM diarization output that merges with Parakeet's word-level transcript to yield a labeled transcript attributing each sentence to the correct speaker.
The analytics layer runs on Nemotron LLM NIM, which extracts structured information from the attributed transcript: revenue guidance, margin commentary, key metric mentions, sentiment by speaker, and deviation from prior quarter language, with JSON output feeding downstream investment analytics systems. The entire pipeline runs on-premises, keeping sensitive investment research data within the institution's infrastructure.
Conclusion
Parakeet TDT v2 NIM, Sortformer, and Nemotron LLM NIM together turn a quarter's worth of earnings calls into attributed, analyzable transcripts within minutes, all on infrastructure you control. Size a single A100 deployment against your call volume first, then consult the NVIDIA NIM catalog for the Parakeet and Nemotron containers.
Links: Sortformer on NVIDIA NGC · NVIDIA NeMo Framework on GitHub · NVIDIA NIM