Speech-to-Text

Code-switching

Two languages in one utterance — error rate and whether the second language survives, per pair, over the live stream.

Request a new language

Tell us which language you want added to this benchmark.

3 models
AssemblyAIuniversal-3-5-proAssemblyAI9.8%3.5%3.7%2.6%4/4
scribe_v2_realtimeElevenLabs75.0%34% both3.8%3.1%7.8%4/4
stt-rt-v5Soniox11.8%4.9%5.1%8.8%4/4
0 → 40% error· underline = best in column

FAQ

What is code-switching in speech recognition?
Code-switching is a speaker mixing two languages inside one utterance. The board scores each model's error rate per language pair over the live stream and whether the second language survives at all; AssemblyAI universal-3-5-pro handled 4 / 4 tested pairs.
Which STT models handle Spanish-English code-switching?
On the board's English-Spanish clips, AssemblyAI universal-3-5-pro measured 3.5% WER, ElevenLabs scribe_v2_realtime 3.8%, and Soniox stt-rt-v5 4.9%.
Which STT models handle Mandarin-English mixed speech?
AssemblyAI universal-3-5-pro measured 9.8% mixed-token error on Mandarin-English and Soniox stt-rt-v5 11.8%; ElevenLabs scribe_v2_realtime anglicizes the Mandarin and measured 75.0%.
Why do STT models fail on code-switched audio?
Most streaming recognizers lock onto one language per stream, so mid-sentence switches come out transcribed phonetically in the wrong language. The failure is pair-specific: ElevenLabs scribe_v2_realtime measured 3.1% on English-German but 75.0% on Mandarin-English.
Which STT model covers the most code-switching language pairs?
Soniox stt-rt-v5 and AssemblyAI universal-3-5-pro each handled 4 / 4 of the tested pairs (Mandarin, Spanish, German, and French, each mixed with English); ElevenLabs scribe_v2_realtime handled 3 / 4.