On the HiKE Korean-English benchmark, Whisper-Large made errors on 26.1% of code-switched speech, against 4.4% on monolingual English and 3.2% on monolingual Korean (Paik et al., HiKE benchmark 2026). That is roughly six times the English error rate (26.1 / 4.4 = 5.9) for the same model. The pattern repeats across language pairs: the CS-FLEURS authors measured character error rates over 2x higher on code-switched recordings than on monolingual FLEURS (Yan et al., Interspeech 2025), and in the ASRU 2019 Mandarin-English challenge the winning system still erred on 12.33% of embedded English words versus 4.04% of Mandarin characters (Shi et al., ASRU 2019 challenge report). Code-switching (moving between two languages inside one conversation, sentence or even word) is ordinary behavior for many of the people described in our bilingualism and multilingual households roundup, yet recognition systems are still tuned and ranked on one language at a time, much like the accent gaps measured in single-language benchmarks. We aggregated data from the SEAME corpus at the Linguistic Data Consortium, the MUCS 2021 and ASRU 2019 challenges, CS-FLEURS, HiKE, CS-Dialogue, ASCEND, the PIER metric paper, the DISPLACE 2023 challenge, Pew Research Center, and ACL survey work on code-switching research.
A note on terms used throughout: the “matrix” language is the main language of a sentence, the “embedded” language is the one that is switched in, and a “switch point” is the exact word where the language changes.
TL;DR
- Whisper-Large: 26.1% mixed error rate on Korean-English code-switched speech vs 4.4% WER on monolingual English (Paik et al., HiKE 2026).
- CER is over 2x higher on code-switched CS-FLEURS sets than on monolingual FLEURS (Yan et al., Interspeech 2025).
- Best ASRU 2019 system: 4.04% Mandarin CER but 12.33% English WER inside the same mixed sentences (Shi et al., ASRU 2019 challenge).
- Error on tokens right after a switch into Yoruba: 0.95 to 1.00 across ASR systems and audio language models (Okocha and Grant, arXiv 2026).
- Fine-tuning improved Whisper-large WER on Spanish-English by 5.06% while switch-point error (PIER) got 39.47% worse (Ugan et al., ICASSP 2025).
- Average language segment in Indian bilingual conversations: around 500 ms, many only about 200 ms (Baghel et al., DISPLACE 2023).
- Whisper-small zero-shot on the SEAME DevSge dev set: 157.5% MER when told to expect Mandarin vs 83.6% when told English (Yang et al., 2023).
- TALCS, the largest listed Mandarin-English corpus, holds 587 hours, about 0.09% of Whisper’s 680,000 training hours (Zhou et al., CS-Dialogue 2025; Radford et al., OpenAI 2022).
- 72 of 114 code-switching corpora introduced in the NLP literature are private (Winata et al., ACL Findings 2023).
- 63% of US Latino adults use Spanglish at least sometimes; 77% of bilingual Latinos do (Pew Research Center 2023).
- Around 3.5% of 50 million tweets in seven languages were code-switched (Rijhwani et al., ACL 2017).
1. The Code-Switching Penalty: Mixed vs Monolingual Error Rates
The headline finding is consistent across every benchmark that runs the same model on both kinds of speech: error rates jump by multiples once two languages share a sentence. On HiKE, Whisper-Large’s code-switched error rate was 5.9 times its English WER (26.1 / 4.4) and 8.2 times its Korean CER (26.1 / 3.2). HiKE is the cleanest comparison because it reports monolingual FLEURS scores for the same models alongside the code-switched results. GPT-4o-transcribe, the strongest system in that table, still went from 3.3% WER on English to 21.8% MER on mixed speech.
The gap is not symmetrical. In mixed sentences the embedded language suffers most: the ASRU 2019 winner made about three times as many errors on English words as on Mandarin characters (12.33 / 4.04 = 3.05), even though English made up only a small part of each sentence: an average of 1.6 English words against 8.6 Chinese characters. In the English-Yoruba evaluation the imbalance is extreme, with near-total failure on the lower-resource language.
| Metric | Value | Source |
|---|---|---|
| Whisper-Large, code-switched Korean-English (mixed error rate) vs monolingual English WER vs monolingual Korean CER | 26.1% vs 4.4% vs 3.2% | Paik et al., HiKE benchmark 2026 |
| GPT-4o-transcribe, code-switched Korean-English (mixed error rate) vs monolingual English WER | 21.8% vs 3.3% | Paik et al., HiKE benchmark 2026 |
| Character error rate on code-switched CS-FLEURS vs original monolingual FLEURS | Over 2x higher | Yan et al., CS-FLEURS, Interspeech 2025 |
| Best ASRU 2019 system (Track 1): Mandarin CER vs English WER in the same code-switched sentences | 4.04% vs 12.33% | Shi et al., ASRU 2019 challenge report (2020) |
| Whisper-large-v3 WER: monolingual English TED talks vs Spanish-English Fisher vs Egyptian Arabic-English ArzEn | 6.07% vs 29.43% vs 52.80% | Ugan et al., PIER paper, ICASSP 2025 |
| English-Yoruba test set: error rate on English tokens vs Yoruba tokens (eight faithful systems) | 33% to 46% vs at least 96.6% | Okocha and Grant, arXiv 2026 |
Context note: the Ugan et al. row compares different datasets (lecture English vs conversational telephone speech), so part of that gap reflects recording conditions, not only switching. The HiKE and CS-FLEURS rows hold the model fixed. Full figures: HiKE paper and CS-FLEURS paper.
2. Mixed Error Rate (MER) Benchmarks: Where the Field Stands
Mixed error rate exists because there is no single “word” across Mandarin and English: MER counts errors per character for character-based languages (Mandarin, Korean) and per word for English, then combines them into one rate. That makes MER the common currency of code-switching papers, but it also hides which language is failing, which is why the better papers print CER and WER separately.
Read across benchmarks, the numbers spread widely: the winning system scored 4.94% MER on the ASRU 2019 Mandarin-English test, while the best baseline on spontaneous ASCEND conversations sat at 27.05%. The same language pair can look nearly solved or badly broken depending on recording style and training volume: ASRU 2019 shipped 200 hours of code-switched training audio, ASCEND only 10.62 hours of spontaneous dialogue. Among off-the-shelf models, SenseVoice-Small reached 6.71% MER on CS-Dialogue against 15.29% for Whisper Large-V2.
| Metric | Value | Source |
|---|---|---|
| Best mixed error rate, ASRU 2019 Mandarin-English, Track 1 (traditional ASR, fixed language model) vs Track 3 (end-to-end ASR) | 4.94% vs 5.91% | Shi et al., ASRU 2019 challenge report (2020) |
| Best baseline on ASCEND spontaneous Hong Kong conversations: CER vs MER | 22.69% vs 27.05% | Lovenia et al., ASCEND corpus (2021) |
| CS-Dialogue test set, pre-trained models (MER): SenseVoice-Small vs Paraformer vs Qwen2-Audio vs Whisper Large-V2 | 6.71% vs 9.65% vs 9.76% vs 15.29% | Zhou et al., CS-Dialogue 2025 |
| CS-Dialogue, Conformer trained from scratch: Mandarin CER vs English WER vs MER | 15.45% vs 32.36% vs 18.91% | Zhou et al., CS-Dialogue 2025 |
| MUCS 2021 end-to-end baseline WER on test set: Hindi-English vs Bengali-English | 25.9% vs 31.0% | Diwan et al., MUCS 2021 challenge (Interspeech 2021) |
Outlier note: the CS-Dialogue Conformer row shows the embedded-language problem again, with English WER more than double Mandarin CER in the same test set. Most recent available data for the ASRU 2019, ASCEND (2021) and MUCS 2021 rows, which have not been rerun with current models by their authors. Sources: ASRU 2019 challenge report, CS-Dialogue, MUCS 2021.
3. Switch-Point Failures: What Overall Error Rates Hide
A transcript can score well overall and still get every switch wrong, because the switched words are a small share of the total. PIER (Point-of-Interest Error Rate) was built to measure accuracy only at the words where the language changes, and the results are uncomfortable: after fine-tuning, Whisper-large improved its overall WER on Spanish-English Fisher by 5.06% relative while its switch-point error rose 39.47%. In other words, the usual leaderboard metric rewarded a model that got worse at the actual code-switching.
Direction matters too. On the English-Yoruba set, systems handled switches back into English reasonably and collapsed on switches into Yoruba. Denser switching raised error rates across the strongest systems, and switches inside a single word (a stem in one language, an affix in another) were the hardest of all. Some error rates exceed 100% because insertions are counted as errors; three generative audio models on the Yoruba set over-generated text, a failure mode covered in our speech recognition hallucination roundup.
| Metric | Value | Source |
|---|---|---|
| Fine-tuned Whisper-large on Spanish-English Fisher: change in overall WER vs change in switch-point error (PIER) | 5.06% better vs 39.47% worse (relative) | Ugan et al., PIER paper, ICASSP 2025 |
| Fine-tuned Whisper-large on Fisher: WER vs PIER | 27.94% vs 50.78% | Ugan et al., PIER paper, ICASSP 2025 |
| Whisper-large on ArzEn Arabic-English: switch-point error between words vs inside words | 49.50% vs 105.71% | Ugan et al., PIER paper, ICASSP 2025 |
| Error on the first token after a switch: English to Yoruba vs Yoruba to English | 0.95 to 1.00 vs 0.40 to 0.53 | Okocha and Grant, arXiv 2026 |
| Mean WER of strongest systems at low vs high switch density (English-Yoruba) | 66.9% vs 71.2% | Okocha and Grant, arXiv 2026 |
| Token error rate, English-dominant vs balanced vs Yoruba-dominant utterances | About 0.42 vs 0.65 vs 0.83 | Okocha and Grant, arXiv 2026 |
Methodology caveat: the English-Yoruba set averages 2.60 switches per utterance (maximum 9) across 9.54 hours of audio from 100 speakers, so it is a stress test, not a typical conversation. Full tables: PIER paper and English-Yoruba switch-aware evaluation.
4. Language Identification at Switch Points
Most multilingual models decide which language they are hearing before they transcribe. On code-switched speech that single decision is wrong by design for part of every sentence. In real Indian bilingual conversations, the average stretch of one language lasted around 500 milliseconds, and many turns were single words of about 200 milliseconds (Baghel et al., DISPLACE 2023 summary). Language diarization (labelling who-speaks-which-language-when, the language counterpart of speaker diarization) struggled accordingly, and confusion between languages was the largest error component.
The same weakness shows up as prompt sensitivity. Telling Whisper which language to expect can swing results by tens of points, and letting it detect the language automatically was worse than fixing it on the English-Yoruba set. Script also matters: on CS-FLEURS, pairs written in two different scripts had on average 3x higher CER than pairs sharing one script, a related problem to the cross-language representation issues covered in our multilingual speech embedding statistics.
| Metric | Value | Source |
|---|---|---|
| DISPLACE 2023 language diarization error rate, Phase-2 evaluation: baseline (B2) vs best team | 41.67% vs 37.60% | Baghel et al., DISPLACE 2023 challenge summary |
| Confusion-error component of the baseline (B2), Phase-2 evaluation | 26.40% (largest component) | Baghel et al., DISPLACE 2023 challenge summary |
| DISPLACE 2023 corpus: total hours and languages | 32 hours; seven languages (six Indian languages plus Indian English) | Baghel et al., DISPLACE 2023 challenge summary |
| Whisper-small zero-shot MER on the SEAME DevSge dev set: prompted for Mandarin vs prompted for English | 157.5% vs 83.6% | Yang et al., Whisper on SEAME and ASRU 2019 (2023) |
| Whisper-large-v3 on English-Yoruba: WER with automatic language detection vs fixed prompt | 74.8% vs 71.2% | Okocha and Grant, arXiv 2026 |
| Whisper-large-v3 switch-point error (PIER) on the SEAME.sge test split: language-agnostic decoding vs English-forced decoding | 64.51% vs 36.10% | Ugan et al., PIER paper, ICASSP 2025 |
| Whisper-Large-v3 CER on read CS-FLEURS speech: same-script pairs vs distinct-script pairs | 7.32% vs 32.33% | Yan et al., CS-FLEURS, Interspeech 2025 |
Context note: the DISPLACE summary also reports a lenient score (DER**, 34.72% for the baseline in Phase 2) that applies a 250 ms collar, which removes short foreign words of 500 ms or less from evaluation. The paper’s authors conclude that the primary error comes from the short duration of the code-mixed stretches. Source: DISPLACE 2023 challenge summary.
5. Data Scarcity: How Many Hours of Code-Switched Speech Exist
The shortage is the root cause behind sections 1 to 4. The largest Mandarin-English corpus in the CS-Dialogue comparison table, TALCS, holds 587 hours; Whisper was trained on 680,000 hours of audio, so TALCS equals about 0.09% of that (587 / 680,000 = 0.086%). Most code-switching corpora are far smaller, cover one language pair, and many cannot be downloaded at all.
Even the benchmark everyone cites is hard to pin down: SEAME appears as 192 hours in the Linguistic Data Consortium catalog but as 30 hours in one 2025 comparison table (see the Data watch note). Vocabulary coverage is a second gap: in MUCS 2021, the out-of-vocabulary rate (share of test words missing from the training vocabulary) reached 27.3% on the Bengali-English blind test set.
| Metric | Value | Source |
|---|---|---|
| SEAME Mandarin-English corpus (Singapore and Malaysia): hours and speakers | About 192 hours; 156 speakers aged 19 to 33 | Linguistic Data Consortium, SEAME LDC2015S04 (2015) |
| MUCS 2021 code-switched training data: Hindi-English vs Bengali-English | 89.86 hours vs 46.11 hours | Diwan et al., MUCS 2021 challenge |
| MUCS 2021 out-of-vocabulary rate, blind test: Hindi-English vs Bengali-English | 19.6% vs 27.3% | Diwan et al., MUCS 2021 challenge |
| ASRU 2019 challenge data: Mandarin-only training vs code-switched training vs code-switched dev | 500 vs 200 vs 40 hours | Shi et al., ASRU 2019 challenge report (2020) |
| Spontaneous Mandarin-English corpora: ASCEND vs CS-Dialogue vs TALCS | 10.62 hours (23 speakers) vs 104.02 hours (200 speakers) vs 587 hours | Zhou et al., CS-Dialogue 2025 (comparison table) |
| CS-FLEURS: languages, code-switched pairs, total hours | 52 languages, 113 pairs, 294 hours | Yan et al., CS-FLEURS, Interspeech 2025 |
| HiKE Korean-English benchmark size | 1,121 utterances, about 2.2 hours, 13 speakers | Paik et al., HiKE benchmark 2026 |
| Code-switching corpora introduced in the NLP literature: total vs private | 114 vs 72 (63%) | Winata et al., ACL Findings 2023 |
Recency flag: the SEAME, MUCS and corpus-survey figures are most recent available data (LDC 2015, MUCS 2021, Winata et al. 2023). Note also that 128 of the 294 CS-FLEURS hours are a training set of synthetic speech generated with text-to-speech, and most of the test sets are synthetic too; only its 17-hour READ test set uses real bilingual voices. Sources: LDC SEAME catalog entry, Winata et al. survey, Whisper paper.
6. Fine-Tuning and Adaptation Results
Small amounts of in-domain code-switched data go a long way. Fine-tuning Whisper-Medium on natural Korean-English speech cut HiKE mixed error rate from 31.3 to 9.0, a drop of 22.3 points, and Whisper-small on SEAME fell from 101.9 to 14.3 MER on the DevMan dev set with the same prompt. The Yang et al. team reported that SEAME gains saturated after roughly 1 to 10 hours of adaptation data.
There is a trade-off, though. The same HiKE fine-tuning raised monolingual Korean CER from 3.4 to 6.0, so a model tuned for mixed speech can lose accuracy on single-language speech. Scale still beats tuning: an in-house Conformer trained on over 40,000 hours outperformed the fine-tuned Whisper-small by about 18% on SEAME (Yang et al., 2023).
| Metric | Value | Source |
|---|---|---|
| Whisper-Medium on HiKE: MER before vs after fine-tuning on natural Korean-English speech | 31.3 vs 9.0 | Paik et al., HiKE benchmark 2026 |
| Same fine-tuning: monolingual Korean CER before vs after | 3.4 vs 6.0 | Paik et al., HiKE benchmark 2026 |
| Whisper-small on SEAME, English prompt: zero-shot vs fine-tuned MER, DevMan dev set and DevSge dev set | 101.9 vs 14.3; 83.6 vs 20.4 | Yang et al., Whisper on SEAME and ASRU 2019 (2023) |
| Adaptation data at which SEAME gains saturated | About 1 to 10 hours | Yang et al., Whisper on SEAME and ASRU 2019 (2023) |
| Whisper-Medium on CS-Dialogue: zero-shot vs fine-tuned MER | 15.88% vs 7.53% (52.6% relative reduction) | Zhou et al., CS-Dialogue 2025 |
| CS-FLEURS read test, seen language pairs: error rate without vs with synthetic code-switched augmentation | 31.77 vs 26.24 | Yan et al., CS-FLEURS, Interspeech 2025 |
Math note: the CS-Dialogue relative reduction is (15.88 - 7.53) / 15.88 = 52.6%. Fine-tuning does not always help: in the PIER paper, Whisper-small fine-tuned on SEAME went from 68.24 to 87.67 MER on the SEAME.sge test split. Source: Yang et al. 2023.
7. How Common Code-Switching Is Among Bilingual Speakers
Recognition systems treat code-switching as an edge case; survey data says it is everyday speech. 63% of US Latino adults say they use Spanglish at least sometimes, and 40% say they do so often (Pew Research Center 2023, survey of 3,029 Hispanic adults conducted August 1 to 14, 2022). Among bilingual Latinos the share rises to 77%, and 52% of them use Spanglish very or extremely often. For a speech recognizer, that means mixed language is the normal input from a large group of users, not a rare exception.
Text data shows the same thing at scale. Across about 50 million tweets in seven languages (Dutch, English, French, German, Portuguese, Spanish and Turkish), around 3.5% were code-switched, with English in 85.4% of them. Research attention follows the biggest bilingual populations: Hindi-English and Spanish-English dominate the literature, and most papers at ISCA, the main speech-research venue, work on speech data.
| Metric | Value | Source |
|---|---|---|
| US Latino adults who use Spanglish at least sometimes vs often | 63% vs 40% | Pew Research Center, Latinos’ Views of and Experiences With the Spanish Language (2023) |
| Spanglish use at least sometimes: bilingual vs English-dominant vs Spanish-dominant Latinos | 77% vs 54% vs 56% | Pew Research Center (2023) |
| Spanglish use at least sometimes: second generation vs third-or-higher generation | 72% vs 45% | Pew Research Center (2023) |
| Bilingual parents who reported mixing languages with their children | 90% of 181 parents | Concordia University, Byers-Heinlein study (2013) |
| Code-switched share of about 50 million tweets in seven languages | Around 3.5% | Rijhwani et al., ACL 2017 |
| Most-mixed pairs as a share of code-switched tweets: English-Spanish, English-French, English-Portuguese | 21.5%, 20.8%, 18.4% | Rijhwani et al., ACL 2017 |
| City with the highest code-switched tweet share vs lowest | Istanbul 12% vs Houston 1% | Rijhwani et al., ACL 2017 |
| Code-switching research papers by language pair: Hindi-English vs Spanish-English vs Chinese-English | 128 vs 86 vs 47 | Winata et al., ACL Findings 2023 |
Recency flag: the parent study (2013) and the Twitter analysis (2017) are most recent available data for those measures; Pew’s figures come from its 2022 survey. Sources: Pew Research Center report, Rijhwani et al. 2017, Concordia study release.
Summary: Code-Switching Speech Recognition by the Numbers
| Metric | Value | Source |
|---|---|---|
| Whisper-Large: code-switched Korean-English MER vs monolingual English WER | 26.1% vs 4.4% | Paik et al., HiKE benchmark 2026 |
| CER on code-switched CS-FLEURS vs monolingual FLEURS | Over 2x higher | Yan et al., CS-FLEURS, Interspeech 2025 |
| ASRU 2019 winner: Mandarin CER vs embedded English WER | 4.04% vs 12.33% | Shi et al., ASRU 2019 challenge report (2020) |
| Best baseline MER on spontaneous ASCEND conversations | 27.05% | Lovenia et al., ASCEND corpus (2021) |
| CS-Dialogue MER: SenseVoice-Small vs Whisper Large-V2 | 6.71% vs 15.29% | Zhou et al., CS-Dialogue 2025 |
| MUCS 2021 end-to-end baseline WER: Hindi-English vs Bengali-English | 25.9% vs 31.0% | Diwan et al., MUCS 2021 challenge |
| Fine-tuned Whisper-large on Fisher: WER change vs switch-point error change | 5.06% better vs 39.47% worse | Ugan et al., PIER paper, ICASSP 2025 |
| Error after switching into Yoruba vs into English | 0.95 to 1.00 vs 0.40 to 0.53 | Okocha and Grant, arXiv 2026 |
| Average language segment in Indian bilingual conversations | Around 500 ms | Baghel et al., DISPLACE 2023 challenge summary |
| Best language diarization error rate, DISPLACE 2023 | 37.60% | Baghel et al., DISPLACE 2023 challenge summary |
| Read CS-FLEURS CER: same-script vs distinct-script pairs | 7.32% vs 32.33% | Yan et al., CS-FLEURS, Interspeech 2025 |
| Largest listed Mandarin-English corpus (TALCS) vs Whisper training data | 587 hours vs 680,000 hours | Zhou et al., CS-Dialogue 2025; Radford et al., OpenAI 2022 |
| SEAME corpus size | About 192 hours, 156 speakers | Linguistic Data Consortium, LDC2015S04 |
| Private share of code-switching corpora | 72 of 114 (63%) | Winata et al., ACL Findings 2023 |
| US Latino adults using Spanglish at least sometimes | 63% | Pew Research Center (2023) |
| Code-switched share of tweets in seven languages | Around 3.5% | Rijhwani et al., ACL 2017 |
Methodology and Sources
- Paik et al., HiKE: Hierarchical Evaluation Framework for Korean-English Code-Switching Speech Recognition, arXiv v4, January 13, 2026: arxiv.org/html/2509.24613v4
- Yan et al., CS-FLEURS: A Massively Multilingual and Code-Switched Speech Dataset, Interspeech 2025: arxiv.org/abs/2509.14161v1
- Okocha and Grant, Beyond Word Error Rate: A Switch Aware Evaluation of ASR and Audio Language Models on English Yoruba Code-Switched Speech, arXiv, September 10, 2026: arxiv.org/abs/2609.11786
- Ugan, Pham, Barmann and Waibel, PIER: A Novel Metric for Evaluating What Matters in Code-Switching, ICASSP 2025: arxiv.org/html/2501.09512
- Zhou et al., CS-Dialogue: A 104-Hour Dataset of Spontaneous Mandarin-English Code-Switching Dialogues for Speech Recognition, arXiv v2, March 12, 2025: arxiv.org/html/2502.18913v2
- Shi, Feng and Xie, The ASRU 2019 Mandarin-English Code-Switching Speech Recognition Challenge: Open Datasets, Tracks, Methods and Results, 2020: arxiv.org/abs/2007.05916
- Diwan et al., Multilingual and code-switching ASR challenges for low resource Indian languages (MUCS 2021), 2021: arxiv.org/abs/2104.00235
- Lovenia et al., ASCEND: A Spontaneous Chinese-English Dataset for Code-switching in Multi-turn Conversation, 2021: arxiv.org/abs/2112.06223v4
- Yang, Peng, Zhong, Huang and Chng, Adapting OpenAI’s Whisper for Speech Recognition on Code-Switch Mandarin-English SEAME and ASRU2019 Datasets, November 29, 2023: arxiv.org/html/2311.17382v1
- Baghel et al., Summary of the DISPLACE Challenge 2023, revised January 3, 2024: arxiv.org/abs/2311.12564
- Linguistic Data Consortium, SEAME catalog entry LDC2015S04, released April 15, 2015: catalog.ldc.upenn.edu/LDC2015S04
- Winata, Aji, Yong and Solorio, The Decades Progress on Code-Switching Research in NLP, ACL Findings 2023: arxiv.org/abs/2212.09660
- Radford et al. (OpenAI), Robust Speech Recognition via Large-Scale Weak Supervision, 2022: arxiv.org/abs/2212.04356
- Pew Research Center, Latinos’ Views of and Experiences With the Spanish Language, September 20, 2023: pewresearch.org
- Rijhwani et al., Estimating Code-Switching on Twitter with a Novel Generalized Word-Level Language Detection Technique, ACL 2017: aclanthology.org/P17-1180
- Concordia University, Byers-Heinlein study of 181 bilingual parents, press release January 16, 2013: sciencedaily.com
- Data watch: SEAME’s size differs by source: the LDC catalog lists about 192 hours and 156 speakers, the CS-Dialogue comparison table lists 30 hours and 157 speakers (its own text says 92 speakers), and Yang et al. used a 93.6-hour training set, so the figures likely describe different subsets. The MUCS 2021 abstract cites a 32.45% code-switching baseline WER that matches the unaligned end-to-end average, not the realigned results quoted here, and its stated total of about 600 hours does not match the sum of its own tables. The ASRU 2019 abstract says 240 hours of code-switched data, which is the 200 training hours plus 40 dev hours. HiKE numbers changed between arXiv versions (Whisper-Tiny overall MER 103.8 in v2, 93.6 in v4); we quote v4. The English-Yoruba paper was posted on September 10, 2026 and its numbers may change after peer review. MER, CER, WER and PIER can exceed 100% because insertions count as errors. Prevalence figures measure different things (self-reported speech, parent-reported speech, written tweets) and should not be combined.
- Last updated: October 31, 2026. We update this roundup quarterly, and the next refresh is expected when the Interspeech 2026 and ICASSP 2027 code-switching papers and the final peer-reviewed version of the HiKE benchmark are published.