Transcriber Models

Base and fine-tuned transcription models

Whisper Large v3

Base Active

English, Lithuanian · Updated Feb 20

WER18.4%

CER5.2%

Latency1.2s

Size1.55 GB

Whisper Large v3 — Fine-tuned v2

Fine-tuned Draft

Lithuanian · Trained on Domain-Specific Lithuanian v1 · Feb 24

WER11.2%

CER3.1%

Latency1.3s

Size1.58 GB

Whisper Medium

Base Archived

English · Updated Jan 10

WER22.1%

CER7.4%

Latency0.8s

Size769 MB

Whisper Large v3 — Fine-tuned v1

Fine-tuned Archived

Lithuanian · Trained on Domain-Specific Lithuanian v1 · Feb 18

WER14.8%

CER4.3%

Latency1.3s

Size1.57 GB

Model Comparison

Compare all available transcription models across key performance metrics.

Model	Type	Language	WER %	CER %	Latency	Size	Status
Whisper Large v3	Base	English, Lithuanian	18.4%	5.2%	1.2s	1.55 GB	Active
Whisper Large v3 — Fine-tuned v2	Fine-tuned	Lithuanian	11.2%	3.1%	1.3s	1.58 GB	Draft
Whisper Medium	Base	English	22.1%	7.4%	0.8s	769 MB	Archived
Whisper Large v3 — Fine-tuned v1	Fine-tuned	Lithuanian	14.8%	4.3%	1.3s	1.57 GB	Archived

WER Benchmarking

Compare model performance on a standard test dataset. Lower WER/CER is better. Latency measured as real-time factor (RTF).

Model	WER %	CER %	Latency (RTF)	Samples	Status
Whisper Large v3 — Fine-tuned v2	11.2%	3.1%	0.42x	50 / 50	Complete
Whisper Large v3 — Fine-tuned v1	14.8%	4.3%	0.43x	50 / 50	Complete
Whisper Large v3 (Base)	18.4%	5.2%	0.38x	50 / 50	Complete
Whisper Medium (Base)	22.1%	7.4%	0.21x	50 / 50	Complete

WER Comparison

FT v2

11.2%

FT v1

14.8%

Large v3

18.4%

Medium

22.1%

Bar width represents relative WER (lower is better). Best model highlighted in green.