Historical ASR benchmark

中文

This is a historical record with incomplete provenance, retained for readers of earlier FunASR comparisons. It is not a new measurement, a universal leaderboard, or a guarantee for a current checkpoint, machine or deployment. For a new evaluation, start with the performance methodology.

Historical Summary

The table below preserves the original report's wording and numbers. Its "best" labels refer only to that report, not to all available models or hardware.

The 169.6x full run and 211.8x initial run are separate reported results. The original page does not disclose the measurement date. The source was checked on 2026-09-07, which is a snapshot audit date, not the measurement date.

Historical Results

All values and notes in this table are archived claims from the original report. The notes are not current API capability guarantees. In particular, a model's raw tags do not imply that an HTTP endpoint returns those tags, and an old timestamp limitation must not replace the current model selection guide.

The repeated CPU/GPU CER values do not establish independent per-device scoring. The raw predictions, references and scoring program are not available in the audited record. The tag-stripping statement is preserved as a historical claim, not as newly verified scorer output. Numerical precision and rounded speed/RTF pairs are retained without recalculation.

Provenance and Limitations

The original English HTML is pinned to its historical GitHub Pages commit. Its bytes matched the archived public page during the source audit. That establishes the source of this table, not the correctness or reproducibility of the underlying measurements.

The old report describes RTF as total inference time divided by total audio duration, and speed as its reciprocal. The latter is also called RTFx:

RTF  = total inference time / total audio duration
RTFx = total audio duration / total inference time = 1 / RTF

The following commands are historical text and cannot be run directly from the audited checkout. All three referenced files were absent at FunASR source revision 386f6f9106684ba5a114e796147db4396a09eab5; no replacement scripts or reproduction data are supplied by this document.

python benchmark/run_full_benchmark.py
python benchmark/run_remaining.py
python benchmark/fix_sensevoice_cer.py

The audited report does not provide a CPU model/thread count, dataset membership and reference manifest, exact checkpoint revisions, software/driver versions, per-file predictions or timing logs, or a complete timing scope covering warmup, I/O and preprocessing. Without these materials, the old table is not directly reproducible and its CPU-versus-GPU headline is not a production-wide guarantee.

This record's 11,539 seconds must remain separate from the 11,541 seconds reported in the vLLM methodology. Both mention 184 files, but that alone does not prove identical dataset membership. Do not merge their rows or silently normalize the two-second discrepancy.

Choosing a Current Path

The following is the original recommendation table, retained as historical context only. It is not a newly validated recommendation or performance ranking.

For current decisions, use the model and capability guide, Agent integration contracts, and vLLM deployment guide. The separate MLT checkpoint's language coverage must not be attributed to Fun-ASR-Nano. Evaluate your own audio, runtime and end-to-end latency before selecting a deployment.

Use the performance methodology for new measurements and the WebSocket benchmark guide for concurrent realtime services. The current migration timing helper measures FunASR on your own audio; it does not compute CER/WER, run Whisper or reproduce the missing historical scripts. Keep your timing scope, failed files and quality evaluation separate and explicit.