Audio transcription metrics (RTF & RTFx)

[AlonKellner-RedHat]

Problem Statement

When benchmarking audio transcription one of the more common metrics that's used is the Real-Time-Factor (RTF), or its inverse RTFx.

Currently, GuideLLM doesn't provide any insight into this when benchmarking audio, while it's very simple to calculate - E2E latency divided by audio duration (RTF) or the other way around (RTFx).

Proposed Solution

When using the audio_transcription or the realtime_transcription request types, automatically report RTF and RTFx.

Alternatives Considered

We could accept that RTF and RTFx won't be available when benchmarking audio transcription.

Usage Examples

python3 -m vllm.entrypoints.openai.api_server \
  --model mistralai/Voxtral-Mini-4B-Realtime-2602 \
  --tokenizer-mode mistral \
  --config-format mistral \
  --load-format mistral \
  --trust-remote-code \
  --compilation-config '{"cudagraph_mode":"PIECEWISE"}' \
  --tensor-parallel-size 1 \
  --max-model-len 45000 \
  --max-num-batched-tokens 8192 \
  --max-num-seqs 16 \
  --gpu-memory-utilization 0.90 \
  --host 0.0.0.0 --port 8000


guidellm benchmark \
  --target http://localhost:8000/v1 \
  --request-type realtime_transcription \
  --backend openai_realtime_ws \
  --data /workspace/custom-audio-dataset/hf_dataset \
  --profile synchronous \
  --max-requests 10 \
  --output-dir /workspace/repo/runs/2026-04-23T13-41-41 \
  --outputs json,html,csv

# RTF & RTFx always reported when `request-type=realtime_transcription` or `request-type=audio_transcription`

Additional Context

#713
https://openvoice-tech.net/index.php/Real-time-factor

I am interested in creating a PR for this 😄