Best local voice transcription model (STT)

The ranking methodology—combining Word Error Rate on common benchmarks with inference speed—remains a sound way to compare local STT models. New whisper forks or alternative ASR models should be evaluated using the same WER and speed metrics to determine if they warrant inclusion in the rankings.

The core criteria of Word Error Rate (accuracy) and inference speed remain the most relevant factors for ranking local STT models. New Whisper forks or alternative ASR approaches can still be fairly compared using these two metrics, so the methodology continues to be sound.

The core criteria of Word Error Rate and inference speed remain the appropriate measures for evaluating local STT models. New whisper forks or alternative ASR models can still be assessed using these same metrics, so the methodology itself is still sound.

The criterion of WER plus inference speed remains appropriate for ranking local STT models, but recent releases such as NVIDIA NeMo ASR, WhisperX, and other optimized Whisper forks (e.g., Whisper.cpp updates, HuggingFace's distil-whisper) have emerged that could shift the rankings. Therefore the current list should be reviewed and updated to include these new contenders.

Recent releases such as faster-whisper updates, NVIDIA NeMo ASR, and open-source Whisper forks (e.g., WhisperX, Whisper.cpp optimizations) have shown improved WER or speed, which could shift the rankings. The current criterion remains appropriate, but the list should be re‑evaluated to incorporate these newer models.

The ranking methodology—combining Word Error Rate on common benchmarks with inference speed—remains a sound way to compare local STT models. New Whisper forks or alternative ASR models can be evaluated using the same WER + speed criteria, so the approach does not need revision.

The methodology—combining Word Error Rate on common benchmarks with inference speed—remains a sound way to compare local STT models. New Whisper forks or alternative ASR models can be assessed using the same two metrics, so the ranking approach does not need revision.

All models failed — review manually

The core methodology (WER + inference speed) remains appropriate, but recent releases such as Whisper.cpp updates, NVIDIA NeMo ASR, WhisperX, and Coqui STT have shown improved WER/speed trade‑offs that are not reflected in the current rankings. Evaluating these newer forks and non‑Whisper models is necessary to ensure the ranking stays up‑to‑date.

The current methodology relies only on WER and inference speed, but recent releases such as Whisper.cpp updates, NVIDIA NeMo ASR, and WhisperX have shown improved accuracy‑speed trade‑offs that could shift rankings. Evaluating these newer forks and non‑Whisper models would ensure the ranking stays representative of the state‑of‑the‑art.