CoSE-E: A Benchmark for Code-switched Speech Evaluation in Enterprise Settings

Shama Gupta; Hoang H Nguyen; Chelsea Huang; Lindsay Devon Brin; Fanny Riols

cs.SDSep 28, 2026

CoSE-E: A Benchmark for Code-switched Speech Evaluation in Enterprise Settings

Authors: Shama Gupta, Hoang H Nguyen, Chelsea Huang, Lindsay Devon Brin, Fanny Riols

Organizations: ServiceNow AI Research · Qualcomm Technologies, Inc.

Abstract

Code-switching (CS), a seamless alternation between languages within a single utterance, remains a critical challenge in automatic speech recognition (ASR). While prior works focus on conversational CS-ASR, enterprise settings demand evaluation of operational impact beyond edit-distance errors: how code-switching transcription errors propagate to downstream voice agent task failures. In this work, we propose (1) a CS-ASR synthetic benchmark and multidimensional evaluation framework tailored to enterprise domains, (2) systematic evaluation of frontier ASR systems across 5 language pairs, (3) diagnostic analysis of the additional transcription errors that code-switching introduces across language pairs and models. We release COSE-E to support enterprise-focused CSASR evaluation for multilingual voice agents in enterprise deployment.

Figures & tables

Figure 1: Overview of CoSE-E Generation Pipeline

Language Pair	Dataset	Matrix Language	Embedded Language	# utts	words/utt	chars/utt	switches/utt	CMI	switch density	% mono ( $\downarrow$ )
ES/EN	CoSE-E	ES (64.2%)	EN (35.8%)	259	24.60	114.76	6.91	0.32	0.29	0.00
	CS-FLEURS	ES (80.7%)	EN (19.2%)	1576	26.01	129.33	5.90	0.16	0.24	3.24
	SWITCHLINGUA	ES (64.5%)	EN (35.5%)	8110	54.95	281.99	6.71	0.32	0.13	0.95
FR/EN	CoSE-E	FR (57.2%)	EN (42.8%)	298	21.56	98.54	4.95	0.33	0.24	0.00
	CoSE-E (CA)	FR-CA (73.5%)	EN (26.5%)	188	20.82	98.53	5.74	0.26	0.29	0.00
	CS-FLEURS	FR (72.2%)	EN (27.7%)	1315	25.46	129.82	6.17	0.19	0.25	3.42

Table 1: Linguistic characteristics of contemporary code-switching benchmarks . CoSE-E demonstrates consistent code-switching patterns across languages (CMI 0.26–0.33, switch density 0.13–0.29) with no monolingual utterances, unlike baseline corpora which contain significant monolingual content (3.2–67.3%). Column abbreviations: # utts = utterance count ; words/utt = average words per utterance ; chars/utt = average characters per utterance ; switches/utt = code-switches per utterance ; CMI = code-mixing index ; switch density = proportion of switching frames ; % mono = percentage of monolingual content . All values reported to two decimal places.

	ES/EN			FR/EN			FR-CA/EN			DE/EN			ZH/EN
Model	WER	SWER	AER	WER	SWER	AER	WER	SWER	AER	WER	SWER	AER	WER	CER	SWER	AER
Scribe-V2	.022	.004	.033	.031	.006	.051	.041	.005	.030	.027	.002	.021	.073	.031	.006	.042
Gemini-3-Flash	.028	.005	.031	.040	.009	.054	.055	.008	.043	.046	.003	.023	.090	.041	.009	.059
AssemblyAI	.029	.004	.033	.039	.009	.062	.052	.006	.034	.048	.003	.023	.093	.046	.010	.057
Qwen3-Omni	.042	.004	.027	.061	.010	.055	.071	.009	.062	.063	.004	.033	.040	.039	.006	.041
Voxtral	.049	.005	.036	.060	.012	.068	.059	.014	.074	.081	.003	.027	–	–	–	–

Table 2: WER, SWER, and AER by model and language pair. ZH/EN WER is jieba word-level (mean of per-utterance WER, as for all pairs) and broadly comparable to the Latin-script pairs; ZH/EN CER is character-level and not magnitude-comparable. All ZH/EN outputs were normalized Traditional

\rightarrow

Simplified (OpenCC) before WER/CER/SWER scoring. Best per column in bold; ties at displayed precision (3 decimals) are both bolded.

Language pair	WER	SWER	AER
EN/DE	2.62	1.38	1.50
EN/ES	1.75	2.00	2.25
EN/FR-CA	3.62	3.13	2.88
EN/FR	2.00	3.50	3.38
Friedman $\chi^{2}(3)$	10.05	13.95	9.45
$p$	.018	.003	.024

Table 3: Mean rank of each language pair per metric (4 pairs, 8 models). Rank 1 = easiest, rank 4 = hardest; averaged across all 8 models. Bold = easiest pair per metric. Friedman

\chi^{2}

and Kendall’s

W

summarise whether pairs differ significantly and how consistently models agree on the ordering.

Figure 2: Mean per-utterance code-switching cost ,

\Delta=m(\text{cs})-m(\text{mono})

; positive values indicate code-switching is harder. Rows: WER, SWER, AER. Columns: the two monolingual baselines. Bars are grouped by model (top/mid tier) and colored by language pair; whiskers are

95\%

bootstrap CIs. Stars mark significance under a two-sided sign-flip permutation test, Holm-corrected within each metric

\times

pair

\times

baseline family (

{}^{*}p{<}.05

{}^{**}p{<}.01

{}^{***}p{<}.001

). Whisper is omitted; its deltas reach

+2.5

(see Figure 4 ).

Baseline	entity err.	no err.	RR	OR
	P(AER worsens)
English	0.44	0.054	8.3	14.0
non-English	0.39	0.052	7.6	11.8

Table 4: Entity-error propagation to task failure . RR = risk ratio; OR = odds ratio. When code-switching introduces a critical-entity error, the chance that the downstream answer worsens rises roughly eightfold, near-identically against both monolingual baselines (per-pair RR

6

–

13

, all

p<10^{-30}

). Whisper excluded; records inner-joined by ID.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 3: The AER pipeline ( Pulikodan et al., 2025 ) . LLM1 generates questions from the reference, LLM2 answers from both the reference and ASR hypothesis, and LLM3 judges answer alignment.

Lang.	Qs	Recs	Unan.	Non-u.	Unan.	Non-u.	Flipped	$\rightarrow$ match	$\rightarrow$ mis.	Flip %
			Exp 1 (Judge)		Exp 2 (LLM2+Judge)			Flip direction
EN/ES	777	259	776 (99.9%)	1	766 (98.6%)	11	3	2	1	0.4%
EN/DE	519	173	519 (100.0%)	0	515 (99.2%)	4	5	3	2	1.0%
EN/FR	894	298	893 (99.9%)	1	886 (99.1%)	8	7	4	3	0.8%
EN/FR CA	564	188	561 (99.5%)	3	559 (99.1%)	5	4	3	1	0.7%

Appendix

Table 5: AER pipeline stochasticity across language pairs. Non-unan. = questions where the 3 trials did not all agree. Flipped = records where the majority verdict changed between Experiment 1 and Experiment 2, isolating the effect of LLM2 resampling beyond judge noise.

Language	Annotator	Passed	Total	Pass Rate	AC1
DE	Annotator 1	518	519	99.8%	0.980
	Annotator 2	518	519	99.8%
	Annotator 3	506	519	97.5%
FR	Annotator 1	859	879	97.7%	0.978
	Annotator 2	891	894	99.7%
	Annotator 3	881	888	99.2%

Appendix

Table 6: Per-annotator validation pass rates and Gwet’s AC1 inter-annotator agreement for the code-switched QA validation task. Each annotator independently judged whether every generated question had an answer grounded in the reference utterance. Pass Rate gives the proportion of questions marked valid by that annotator; AC1 is computed once per language across all annotators jointly.

Language	Male Voice	Female Voice	Model	Format
English (monolingual)	Adam	Matilda	V2	24 kHz, 16-bit PCM
German	Finn	Johanna	V2	24 kHz, 16-bit PCM
French Canadian	Felix Tabarnak	Amelie	V2	24 kHz, 16-bit PCM
French	Denis Landrieu	Marine	V2	24 kHz, 16-bit PCM
Spanish	Rodrigo	Cristina Campos	V2	24 kHz, 16-bit PCM
Mandarin Chinese	Jing	Macy	V3	24 kHz, 16-bit PCM

Appendix

Table 7: TTS voices used in the benchmark. Utterances were alternated between male and female voices to achieve 50/50 gender balance. All audio was synthesized at 24 kHz, 16-bit PCM with stability=0.5, similarity_boost=0.75, style=0.2, and speaker boost enabled.

Provider	Model ID	Language-ID Setting	Decoding Parameters
AssemblyAI / Universal-3-Pro	universal-3.5-pro	language_detection: True	defaults, streaming: false
ElevenLabs / Scribe-V2	scribe_v2	language_code omitted	defaults
OpenAI / Whisper-Large-V3-Turbo	whisper-large-v3-turbo	language omitted	translate: false, defaults
Mistral / Voxtral-Small-24B	Voxtral Small 1.0 (24B)	language omitted	temperature: 0
Deepgram / Nova-3-Multilang	nova-3-multilang	language: "multi"	smart_format: true
NVIDIA / Parakeet-TDT-0.6B-V3	parakeet-tdt-0.6b-v3	N/A	defaults

Appendix

Table 8: ASR systems and decoding parameters. All models were run with no language-ID passed.

Metric	$\chi^{2}(7)$	$N$	$p$	Within $W$
WER	1464.1	918	$<.001$	0.23
SWER	612.2	841	$<.001$	0.10
AER	200.1	918	$<.001$	0.03

Appendix

Table 9: Omnibus and concordance statistics on the European language pairs. Friedman tests confirm that the eight models differ significantly on every metric; the reduced

N

for SWER reflects complete-case analysis after judge-failure drops. Within-metric Kendall’s

W

measures how consistently clips agree on model ordering, and declines monotonically from surface form to answer-equivalence (WER

\rightarrow

SWER

\rightarrow

AER), indicating that models trade places more freely as metrics move toward semantics. Cross-metric agreement across the three metric orderings is high (Cross Metric Kendall’s

W=0.963

, 95% CI

[0.889,0.979]

Condition	WER	SWER	AER
Before T $\to$ S normalisation	0.167	0.010	0.059
After T $\to$ S normalisation	0.090	0.010	0.059
$\Delta$	$-46\%$	—	—

Appendix

Table 10: Effect of Traditional

\to

Simplified (OpenCC) normalisation on Gemini-3-Flash(en_zh,

N{=}294

). WER drops 46% once script variants are collapsed before scoring. SWER and AER are unchanged because they judge semantic equivalence, not surface form.

Figure 4: Mean per-utterance code-switching cost for Whisper-only

Language pair	WER	SWER	AER
EN/DE	2.80	1.60	1.60
EN/ES	1.20	1.60	1.80
EN/FR	2.20	4.00	4.40
EN/FR-CA	3.80	3.20	3.00
EN/ZH	5.00	4.60	4.20
Friedman $\chi^{2}(4)$	17.12	15.04	13.60

Appendix

Table 11: Mean rank of each language pair per metric, including ZH/EN (5 pairs, 5 models). Rank 1 = easiest, rank 5 = hardest; averaged across the 5 models with ZH/EN coverage. Bold = easiest pair per metric; SWER ties DE/EN and ES/EN. ZH/EN ranks last on every metric. WER for ZH/EN uses jieba word-level segmentation; see text for comparability caveats.

FR/EN (10/14, 13/14, 4/14 significant)
	WER		SWER		AER
Model	EN	non-EN	EN	non-EN	EN	non-EN
AssemblyAI	$+0.019$ ***	$+0.015$ **	$+0.008$ ***	$+0.005$ *	$+0.042$ **	$+0.018$
Scribe-v2	$+0.023$ ***	$+0.017$	$+0.005$ ***	$+0.003$ **	$+0.022$	$+0.018$
Gemini-3-Flash	$+0.027$ ***	$-0.001$	$+0.007$ ***	$+0.007$ **	$+0.022$	$+0.010$
Qwen3-Omni	$+0.031$ ***	$+0.029$ ***	$+0.010$ ***	$+0.008$ ***	$+0.028$ *	$+0.015$

Appendix

Table 12: Per-model code-switching deltas by language pair, metric, and baseline. Each cell shows the mean per-utterance

\Delta=m(\text{cs})-m(\text{mono})

; positive values indicate code-switching raised error. Significance: *

p{<}.05

, **

p{<}.01

, ***

p{<}.001

(two-sided sign-flip permutation, Holm-corrected within each metric

\times

pair

\times

baseline family). Parenthesized counts in each panel header show significant cells out of total across both baselines. Whisper excluded (see Appendix D.3 ).

Panel 1: CMI
	ES/EN	FR/EN	FR-CA/EN	DE/EN	ZH/EN
Whisper	1.010	0.975	0.949 *	0.969	0.950
Scribe v2	1.028 *	1.008	1.012	1.002	0.994
Nova-3	1.010	0.991	1.007	0.981	–
Parakeet	1.026	1.010	1.000	1.002	–
Voxtral-24B	1.022	0.991	1.032	1.014	–

Appendix

Table 13: Part A odds ratios from per-model, per-language-pair logistic regressions of error occurrence (

\text{WER}>0

) on CMI, switch count, and

\log(n_{\text{words}})

. Stars denote significance: *

p<.05

; **

p<.01

; ***

p<.001

\dagger

AssemblyAI Universal-3.5-Pro. Dashes mark unsupported language pairs.

Language pair	Utterances	$n_{\text{words}}$	Switch count	CMI
		(mean $\pm$ sd, range)	(mean $\pm$ sd, range)	(mean $\pm$ sd, range)
ES/EN	259	$24.6\pm 7.7$ (10–45)	$6.91\pm 3.41$ (1–18)	$31.7\pm 10.5$ (9.4–50.0)
FR/EN	298	$21.6\pm 6.8$ (8–41)	$4.95\pm 2.83$ (1–14)	$33.5\pm 10.4$ (10.5–50.0)
FR-CA/EN	188	$20.8\pm 7.0$ (9–42)	$5.74\pm 2.87$ (1–15)	$25.8\pm 10.1$ (4.0–50.0)
DE/EN	173	$21.1\pm 6.7$ (9–40)	$4.99\pm 2.61$ (1–15)	$30.0\pm 11.2$ (5.4–50.0)
ZH/EN	294	$21.2\pm 6.3$ (12–39)	$2.86\pm 1.59$ (1–12)	$10.3\pm 7.9$ (2.6–50.0)

Appendix

Table 14: Dataset composition per language pair. ZH/EN embeds markedly fewer switches and lower CMI than the other pairs at comparable utterance length.

Model	ES/EN	FR/EN	FR-CA/EN	DE/EN	ZH/EN
Whisper	57.1	82.2	88.3	96.5	96.3
Scribe v2	37.8	47.3	56.4	37.6	31.6
Nova-3	52.9	61.4	80.9	69.9	–
Parakeet	76.8	65.1	81.4	63.0	–
Voxtral-24B	41.3	58.4	70.2	68.8	–
Gemini-3	36.3	50.0	68.6	53.2	54.4

Appendix

Table 15: Error base rate per model

\times

language pair (% of utterances with

\text{WER}>0

). Values near 96–97% indicate a ceiling that limits Part A’s power to detect predictor effects.

\dagger

Universal-3.5-Pro. Dashes mark unsupported pairs.

Language pair	$r_{\text{sw},n}$	$r_{\text{sw},\text{CMI}}$	$r_{n,\text{CMI}}$	max VIF
ES/EN	0.66	$-0.04$	$-0.17$	1.86
FR/EN	0.56	$-0.06$	$-0.06$	1.46
FR-CA/EN	0.65	$0.19$	$-0.20$	2.13
DE/EN	0.60	$-0.13$	$-0.18$	1.60
ZH/EN	0.28	$0.58$	$-0.05$	1.76

Appendix

Table 16: Predictor collinearity among switch count (sw),

\log n_{\text{words}}

(

n

), and CMI. All VIFs fall well below the conventional threshold of 5. ZH/EN is the sole pair where CMI and switch count overlap substantially (

r=0.58

Reference	Je peux accéder à host-2854.corp.local sur mon nouveau laptop. Par contre, I need to set up le VPN pour host-1375.corp.local —vous pouvez me walk through ca?
CS hyp.	(Gemini-3-Flash) host-de-win-54.corp.local … host-365.corp.local AER = 1.0
	(Parakeet) ostewin54.corp.local … OS365.corp.local AER = 1.0
	(AssemblyAI) hostwin54.corp.local … host365.corp.local AER = 1.0
Mono hyp.	(English) host-2854.corp.local … host-1375.corp.local AER = 0.0
	(French) host-2854.corp.local … host-1375.corp.local AER = 0.0

Appendix

Table 17: Destructive entity error: every code-switched hypothesis corrupts at least one hostname (AER = 1.0), while both monolingual hypotheses recover them exactly (AER = 0.0).

Reference	…la query sur sn_safe_story …l’opération query_match sur sn_safe_story
CS hyp.	(AssemblyAI) SN_Safestory … query match … SN_Safe AER = 0.0
	(Parakeet) SNSafeStory … QueryMatch … SnSafeSto AER = 0.0
Mono hyp.	(English) SN_SAFE_STORY … query_match AER = 0.0
	(French) …la requête sur sn_safe_story … query_match AER = 0.0

Appendix

Table 18: Cosmetic entity error: casing and separator changes yield several word-level entity-WER errors, but the values stay recoverable, so AER remains 0.0 across all conditions.

	Entity WER				$\Delta$ vs.
Category	CS	EN	non-EN	$N_{\text{CS}}$	EN	non-EN
IDs	.158	.038	.136	2,023	$+$ .120	$+$ .022
Email	.293	.183	.269	651	$+$ .111	$+$ .025
URL	.160	.050	.187	3,094	$+$ .110	$-$ .027
Phone	.374	.286	.333	91	$+$ .088	$+$ .040
Address	.310	.226	.243	371	$+$ .084	$+$ .067

Appendix

Table 19: Entity WER by category. Against the English baseline, code-switching raises error on IDs, emails, URLs, and phones (

+0.09

–

0.12

); against the non-English baseline the overall effect inverts (

-0.02

), with only small residual increases on IDs and email. Names are unaffected in either comparison.

N_{\text{CS}}

: entity word count in the code-switched condition. Date and measurement counts are not comparable across baselines due to localization and are omitted from the non-English column. Whisper excluded; micro-averaged across models and language pairs.

Baseline	Pair	Worse	Flat	Worse	Flat	$N$	P ${}_{\text{EE}}$
		Entity Error		No Entity Error
EN	ES/EN	93	152	87	1,441	1,773	.380
	FR/EN	157	159	106	1,628	2,050	.497
	FR-CA/EN	92	106	61	1,038	1,297	.465
	DE/EN	38	61	38	1,049	1,186	.384
non-EN	ES/EN	70	136	63	1,520	1,789	.340

Appendix

Table 20: Per-pair entity-error propagation. P

{}_{\text{EE}}

is the probability that AER worsens given an entity error was introduced by code-switching; the no-error base rate is

\sim

0.05 throughout. All per-pair

\chi^{2}

tests yield

p<10^{-30}

. Whisper excluded; records inner-joined by ID.

Explore similar work

Jun 20, 2026cs.CL

Adding Robust Code-Switching Capabilities to High Performance Multilingual ASR

Code-switching (CSW) remains challenging for large multi-lingual ASR systems in real-world deployment. While fine-tuning on synthetic CSW data is possible, it generally degrades strong monolingual baselines. Our goal is to preserve these capabilities while extending models to handle complex code-switching, including morphological variations across languages. We propose Bayesian factorized adaptation, which learns to efficiently integrate switching-relevant knowledge into strong pretrained models without overwriting existing capabilities. Requiring only a small amount of synthetic data, our approach reduces transcription errors by 32.87% on code-switched words while improving overall WER by 5.31%, all while maintaining mono-lingual performance. Our results demonstrate that effective CSW adaptation depends more on knowledge integration than data complexity.

Enes Yavuz Ugan, Alexander Waibel

Interactive Systems Lab, Karlsruhe Institute of Technology (KIT), Germany · InterACT, Carnegie Mellon University (CMU), USA

Code-Switching Multilingual Automatic Speech Recognition Transcription Rapid Adaptation

Ask PDF

May 18, 2026cs.CL

Benchmarking Commercial ASR Systems on Code-Switching Speech: Arabic, Persian, and German

Code-switching -- the natural alternation between two languages within a single utterance -- remains one of the most challenging and under-studied conditions for automatic speech recognition (ASR). We present a benchmark evaluating five commercial ASR providers across four language pairs: Egyptian Arabic--English, Saudi Arabic (Najdi/Hijazi)--English, Persian (Farsi)--English, and German--English, comprising 300 samples per pair selected by a two-stage pipeline combining heuristic filtering with a GPT-4o and Gemini 1.5 Pro ensemble scorer, reducing LLM costs by

\approx

91%. We evaluate on both WER and BERTScore, showing that while both metrics agree on the ordinal ranking of systems for all Arabic and Persian pairs (

τ= 1.0

), WER inflates the magnitude of quality gaps by approximately 3

\times

by penalising semantically correct transliteration choices. ElevenLabs Scribe v2 achieves the lowest WER (13.2% overall) and leads on BERTScore (0.936 overall). Difficulty-stratified analysis reveals performance gaps masked by aggregate averages, and BERT embedding projections confirm semantic proximity between reference and hypothesis despite surface-level script differences. The dataset is publicly available at https://huggingface.co/datasets/Perle-ai/ASR_Code_Switch.

Sajjad Abdoli, Ghassan Al-Sumaidaee, Clayton W. Taylor +2

Perle AI

Automatic Speech Recognition Evaluation Code-Switching Automatic Speech Recognition Multilingual Benchmark Arabic Language Pairs

Ask PDF

Jun 14, 2026cs.SD

Improving Code-Switching ASR with Code-Mixing Guided Synthetic Speech

Code-switch (CS) Automatic Speech Recognition (ASR) remains challenging due to limited availability of high quality CS text-speech pairs for training. Although synthetic data augmentation via Text-to-speech (TTS) has been explored, existing CS TTS approaches primarily optimise reconstruction fidelity and do not explicitly enforce language-boundary consistency, thereby limiting their effectiveness for CS ASR augmentation. This paper proposes a code-mixing guided preference-learning framework that steers synthetic speech generation toward improved code-switching fidelity using the Code Mixing Index (CMI). Experiments on the SEAME Mandarin-English conversational corpus demonstrate that the proposed method enhances the utility of synthetic data for ASR fine-tuning. Specifically, when fine-tuning Whisper Large, the proposed approach reduces Mixed Error Rate (MER) from 12.1%/17.8% to 8.9%/14.2% on the DevMAN and DevSGE sets, respectively.

Yue Heng Yeo, Haoyang Li, Yizhou Peng +6

College of Computing and Data Science, Nanyang Technological University, Singapore · HLT-COE & CLSP, Johns Hopkins University, USA · Institute for Infocomm Research (I2R), A⋆STAR, Singapore +1

Code-Switching Automatic Speech Recognition Preference Learning

Ask PDF