Salad Transcription Accuracy Benchmarks

See how Salad Transcription API performs across multiple languages and datasets.

Accuracy Benchmarks

Salad Transcription API delivers industry-leading accuracy across a wide range of languages and public benchmark datasets. Below is a breakdown of results by language and dataset.

---

Languages with Accuracy ≥ 90%

• English
• Portuguese
• French
• Spanish
• German
• Italian
• Russian

---

Languages with Accuracy between 80%–89%

• Hindi
• Hebrew

---

Languages with Accuracy < 80%

• Urdu
• Kazakh
• Thai (in progress)

---

English

Dataset	Sub-Dataset	Accuracy (Full)	WER (Full)	Accuracy (Lite)	WER (Lite)	Source
TED-LIUM	tedlium	95.8%	4.2%	91.8%	8.2%	TED-LIUM on Hugging Face
Meanwhile	Meanwhile	95.7%	4.3%	83.3%	16.7%	Meanwhile on Hugging Face
CommonVoice	cv-corpus-5.1-2020-06-22	95.1%	4.9%	81.3%	18.7%	Common Voice
CommonVoice	cv-corpus-20.0-delta-2024-12-06	93.1%	6.9%	78.1%	21.9%	Common Voice

---

Portuguese

Dataset	Sub-Dataset	Accuracy	WER	Source
CommonVoice	cv-corpus-8.0-2022-01-19	92.0%	8.0%	Common Voice

---

French

Dataset	Sub-Dataset	Accuracy	WER	Source
CommonVoice	cv-corpus-10.0-delta-2022-07-04	92.0%	8.0%	Common Voice

---

Spanish

Dataset	Sub-Dataset	Accuracy	WER	Source
CommonVoice	cv-corpus-12.0-delta-2022-12-07	94.0%	6.0%	Common Voice
CommonVoice	cv-corpus-14.0-delta-2023-06-23	96.8%	3.2%	Common Voice
CommonVoice	cv-corpus-16.1-delta-2023-12-06	96.8%	4.3%	Common Voice

---

German

Dataset	Sub-Dataset	Accuracy	WER	Source
CommonVoice	cv-corpus-13.0-delta-2023-03-09	96.3%	3.7%	Common Voice

---

Hindi

Dataset	Sub-Dataset	Accuracy	WER	Source
CommonVoice	cv-corpus-20.0-2024-12-06	84.0%	16.0%	Common Voice

---

Italian

Dataset	Sub-Dataset	Accuracy	WER	Source
CommonVoice	—	93.3%	6.7%	Common Voice

---

Russian

Dataset	Sub-Dataset	Accuracy	WER	Source
CommonVoice	—	96.4%	3.6%	Common Voice

---

Hebrew

Dataset	Sub-Dataset	Accuracy	WER	Source
CommonVoice	cv-corpus-17.0-2024-03-15	84.2%	15.8%	Common Voice

---

Kazakh

Dataset	Sub-Dataset	Accuracy	WER	Source
CommonVoice	cv-corpus-19.0-2024-09-13	51.0%	49.0%	Common Voice

---

Urdu

Dataset	Sub-Dataset	Accuracy	WER	Source
CommonVoice	cv-corpus-9.0-2022-04-27	78.8%	21.2%	Common Voice

---

Thai (in progress)

Dataset	Sub-Dataset	Accuracy	WER	Source
CommonVoice	cv-corpus-10.0-delta-2022-07-04	33.0%	67.0%*	Common Voice

Thai WER may need a recalculation due to formatting issues.

---

Methodology

To ensure fair and repeatable accuracy evaluation, we adopted a benchmarking methodology similar to AssemblyAI:

• Public datasets were used for transparency and reproducibility
• Transcripts were normalized using Whisper Normalizer
• Accuracy was calculated using Word Error Rate (WER) via the JiWER library

This benchmark continues to expand as we test more languages and improve our models.

Want to run your own benchmarks? Reach out to us at [e-mail].