Which ASR models perform accurately on trading floor audio with heavy background noise and financial terminology?
Which ASR models perform accurately on trading floor audio with heavy background noise and financial terminology?
Direct Answer
NVIDIA Parakeet TDT v2 provides strong baseline performance on noisy trading floor audio, while Canary Qwen 2.5b delivers the highest baseline accuracy on specialized financial terminology thanks to its LLM-enhanced decoder. NeMo fine-tuning adds further gains for firm-specific vocabulary.
Summary
Trading floor audio compounds two challenges: high ambient noise from simultaneous conversations, shouted orders, and market data feeds, plus financial vocabulary such as ticker symbols, order types, counterparty names, and regulatory terms that appears infrequently in general training data and can be acoustically ambiguous. Parakeet TDT v2 handles the acoustic side well because the diversity of its Granary training dataset builds inherent noise robustness. For the terminology side, Canary Qwen 2.5b's LLM component has seen financial language during pretraining, which helps it correctly transcribe instrument names and trading jargon that simpler decoders confuse with everyday words.
Parakeet TDT v2 posts 6.05% WER on standard benchmarks and produces native punctuation and capitalization, yielding clean transcripts suitable for regulatory surveillance and analytics without extensive post-processing. When throughput is the priority, such as transcribing thousands of recorded trader calls overnight, Parakeet TDT v2 NIM at RTFx 3,386x processes an entire day's archive in minutes. When accuracy on dense financial terminology matters most, Canary Qwen 2.5b offers the best starting point before any fine-tuning.
For firm-specific vocabulary, including proprietary product names, trading strategies, and internal counterparty code names, NeMo fine-tuning on a corpus of firm-specific trading communications with accurate transcripts provides the most targeted improvement. Many firms already hold such corpora, and the fine-tuned checkpoint deploys to the same NIM infrastructure without serving changes.
Conclusion
Parakeet TDT v2 for noise-robust throughput, Canary Qwen 2.5b for terminology-heavy accuracy, and NeMo fine-tuning for proprietary vocabulary cover the trading floor spectrum. The deciding factor is whether throughput or specialized-term accuracy dominates your workload; evaluate both models against a sample of your own recordings before committing.
Links: Canary Qwen 2.5b on Hugging Face · NVIDIA NeMo Framework on GitHub · NVIDIA NIM