Bilingual Language Identification

• Bilingual Language Identification (BLID) is a method for detecting and distinguishing two languages within texts, speech, and handwriting even when language switches occur arbitrarily.
• It leverages approaches from character-level BiRNNs and convolutional models to transformer-based speech predictors, addressing span-level, word-level, and frame-synchronous tasks.
• Research in BLID highlights persistent challenges such as accented speech, low-resource scenarios, and code-mixing, guiding developments in confusion resolution and context-aware adaptation.

Bilingual Language Identification (BLID) is the task of identifying and distinguishing between two languages in a given text, and in multilingual speech systems it also includes inferring language on-the-fly when language identity is not known beforehand (Yigezu et al., 5 Sep 2025, Punjabi et al., 2020). Although often treated as a special case of Language Identification (LID), BLID is materially harder than ordinary monolingual classification because languages may switch arbitrarily within a document, multilingual speakers introduce accent- and locale-dependent confusability, and many deployed LID systems remain optimized for clean monolingual inputs rather than bilingual or code-switched data (Kocmi et al., 2017, Titus et al., 2020, Ojo et al., 22 Sep 2025).

1. Task scope and granularity

Across the literature, BLID is defined at multiple granularities. In multilingual text identification, a document may contain two or three languages and the immediate goal may be to return the language set; a finer formulation labels spans when languages change arbitrarily within a document; and still finer settings operate at the word or character level (Kocmi et al., 2017, Mandal et al., 2018). In speech, BLID appears as utterance-level classification, frame-synchronous streaming prediction, chunk aggregation over fixed windows, or turn-level assignment after segmentation by speaker diarization (Tjandra et al., 2021, Zhang et al., 2022, Valente et al., 2024). BLID also appears in script-discrimination settings, such as identifying Arabic words inside Arabic/Latin bilingual handwriting by exploiting structural character features rather than global statistics (Haboubi et al., 2011).

Modality	Typical decision unit	Representative formulation
Text	document, word, character span	return both languages present; identify contiguous spans
Speech	chunk, utterance, frame, speaker turn	aggregate chunk predictions; per-frame streaming LID; turn-level labeling
Handwriting/script	word image	Arabic vs. Latin discrimination from structural primitives

This diversity of granularity is not incidental. It reflects different downstream requirements: spell-check and short-string routing emphasize extremely short text; code-switched NLP requires span or word labels; bilingual ASR requires low-latency utterance or frame decisions; and multilingual broadcast transcription benefits when language changes align with speaker turns. A plausible implication is that BLID is better viewed as a family of closely related inference problems than as a single classification task.

2. Textual, span-level, and script-aware methods

Character-level sequence models constitute one major BLID lineage. LanideNN processes windows of 200 Unicode characters with 200-dimensional character embeddings and a BiRNN with 500 GRU cells in each direction, producing a probability distribution over language labels for each character and then grouping contiguous stretches of the same label into spans (Kocmi et al., 2017). On short texts of about 142 characters, it reports 95.0% accuracy over 131 languages, and on the ALTW2010 bilingual shared task it reports a micro-averaged $F_1$ of 0.965, or 0.941 without retraining on specific in-domain data (Kocmi et al., 2017). This formulation is explicitly designed for multilingual inputs where switches do not align with token boundaries.

Word-level BLID for code-mixed text has also been approached with hybrid architectures. A multichannel neural network for Bengali-English and Hindi-English code-mixed data uses four parallel channels over character embeddings: three Conv1D channels with Dropout and MaxPooling, and one two-layer LSTM channel, followed by dense layers that output a sigmoid probability for the word label (Mandal et al., 2018). A downstream Bi-LSTM-CRF context module then consumes the word model’s confidence score together with a character embedding, allowing contextual disambiguation of ambiguous tokens. The reported accuracies are 93.28% and 93.32% on the two testing sets (Mandal et al., 2018).

A different line of work treats tokenization itself as the BLID mechanism. UniLID learns language-conditional unigram distributions over a shared tokenizer vocabulary and treats segmentation as a language-specific phenomenon, yielding a closed-set decision rule

$$\hat{\ell}(s) = \argmax_{\ell \in \mathcal{L}} p(\ell \mid s), \qquad p(\ell \mid s) = \frac{p(s \mid \ell)}{\sum_{\ell' \in \mathcal{L}} p(s \mid \ell')}.$$

At inference, it approximates $p(s \mid \ell)$ with the most likely segmentation under the language-specific UnigramLM, found by a Viterbi-style dynamic program (Meister et al., 19 Feb 2026). On WiLI, it surpasses 70% accuracy with as few as five labeled samples per language and exceeds 92% accuracy with 50 samples per language; on DSL-ML 2024 it improves macro $F_1$ from 0.53 for fastText to 0.72 (Meister et al., 19 Feb 2026). This is especially relevant to BLID when the two candidate languages are closely related or dialectal.

Hierarchical error-resolution models target a different failure mode: systematic confusion. LIMIT begins with a high-coverage root model, inspects its empirical confusion matrix on annotated data, and attaches small confusion-resolution units to clusters of highly confused languages or triplets (Agarwal et al., 2023). Each such unit is a lightweight fastText classifier trained on as little as 1000 sentences per language. On MCS-350, a corpus of 52K children’s stories spanning 350+ languages, LIMIT improves average $F_1$ from 0.29 to 0.68 in confusion clusters, and on FLORES-200 it improves $F_1$ from 0.77 to 0.86 (Agarwal et al., 2023). Because BLID is often dominated by a small number of confusion pairs, this cluster-driven strategy is directly aligned with bilingual disambiguation.

Low-resource lexical BLID has also been studied in highly similar languages. For Wolaita and Gofa, an open-source word-level dataset of 144,000 words is annotated with three labels—wal, gof, and wal-gof—and the best reported model combines BERT-base-uncased embeddings with an LSTM of 128 units, batch normalization, dense layers of size 768, ReLU activations, dropout of 0.1, and a Softmax output over the three classes (Yigezu et al., 5 Sep 2025). The resulting test $F_1$ is 0.72 (Yigezu et al., 5 Sep 2025). The explicit wal-gof class is notable because it avoids forcing a hard distinction when words are genuinely common to both languages.

Script-aware BLID predates most neural work. In Arabic/Latin handwritten bilingual text, structural primitives such as poles, jambs, upper and lower diacritical dots, loops, PAWs, and letter positions are extracted after baseline estimation and morphological expansion (Haboubi et al., 2011). The method is rule-based rather than classifier-driven, and the authors argue that the distinct distribution of these features can visually discriminate Arabic from Latin. This suggests that, in mixed-script BLID, explicit script structure can remain competitive or complementary to end-to-end modeling.

3. Spoken BLID and joint ASR-LID systems

Recent speech BLID work is strongly shaped by self-supervised pre-training. A wav2vec 2.0-based LID model replaces the standard seven convolutional layers with a time-stacking layer and a linear layer, reducing sequence length by a factor of 4, then applies a context encoder consisting of a linear layer, layer normalization, a convolutional layer for relative positions, 24 Transformer layers with hidden size 1024 and 16 heads, and a final linear layer (Tjandra et al., 2021). Inputs are 80-dimensional log-mel spectrogram features from raw audio with a 25 ms window and 10 ms stride. Pre-training combines contrastive and diversity losses,

$$\mathcal{L} = \mathcal{L}_m + \lambda \mathcal{L}_d, \qquad \lambda = 0.1,$$

with vector quantization implemented by Gumbel softmax using 320 codebooks and 2 groups (Tjandra et al., 2021). In a 26-language setup, the cross-lingual XLSR model reaches 89.2% accuracy with only 10 minutes of labeled data per language, 92.8% with 1 hour, 94.2% with 10 hours, and 95.7% with 100 hours; mean pooling performs best, and middle or upper Transformer layers permit pruning from 300M to 100M parameters with little accuracy loss (Tjandra et al., 2021). The paper explicitly notes that, in BLID scenarios, the 26-way task is much harder than BLID, suggesting very high accuracy when only two languages must be separated.

Streaming bilingual ASR has motivated joint BLID models rather than separate LID front-ends. One approach augments the RNN-T output vocabulary with special language tags such as <en>, <es>, and <hi>, appended as utterance-final tokens during training, while also injecting pretrained acoustic-only LID embeddings either into the encoder input, the joint network, or both (Punjabi et al., 2020). To prevent premature tag emission, language tags are penalized during decoding unless their posterior crosses a threshold. On English-Spanish, the bilingual joint ASR-LID architecture matches monolingual ASR and acoustic-only LID accuracies; on English-Hindi, English ASR and LID degrade because code-switching and script ambiguity are more severe (Punjabi et al., 2020).

A larger-scale streaming design modifies a cascaded-encoder RNN-T by adding a per-frame LID predictor. It exploits the difference between a first-pass causal encoder with no right-context and a second-pass right-context encoder with 15 frames of lookahead, and it uses streaming statistics pooling to accumulate running means and standard deviations over time (Zhang et al., 2022). The combined system reaches an average of 96.2% LID prediction accuracy across 9 language locales and achieves the same second-pass WER as using oracle LID as input (Zhang et al., 2022). The architecture is especially relevant to BLID in voice-search and assistant settings because it produces frame-synchronous predictions with little extra test-time cost.

Not all speech BLID requires frame-level switching models. In multilingual broadcast and institutional speech, language changes are reported to be mostly associated with speaker changes, which motivates a cascaded system of speaker diarization followed by segment-based spoken language identification (Valente et al., 2024). Relative to VAD-based segmentation, the proposed system yields up to 10% relative language diarization error reduction, 60% relative language confusion reduction, and more than 8% relative WER reduction on multilingual test sets, while increasing monolingual WER by only 0.1% to 0.7% absolute (Valente et al., 2024). This result is important because it restricts the BLID problem to a turn-level regime when conversational structure supports that assumption.

4. Context, accented speech, and low-resource adaptation

BLID performance for multilingual speakers can be dominated by interaction context rather than acoustics alone. A two-stage system for multilingual dictation first trains a CNN-based acoustic model on coarse-grained targets rather than fine-grained locales, then combines its outputs with device context—installed dictation locales, the currently selected locale, and whether the user toggled locale prior to speaking—through a naïve Bayes model,

$$P(\text{locale} \mid \text{features}) \propto P(\text{locale}) \prod_i P(\text{feature}_i \mid \text{locale}).$$

The acoustic model uses 40-dimensional Mel-filter bank features, a 4-layer CNN frontend with 128 filters and $15 \times 4$ kernels, mean and standard deviation pooling, and about 8 million parameters (Titus et al., 2020). Grouping English into native (L1), non-native (L2), and Indian English (en-IN) improves difficult accent pairs, and the final context-aware system reaches 97.0% Average User Accuracy with 75.1% worst-case tuple accuracy, compared with 92.5% and 66.1% for the best purely acoustic variant in the same table (Titus et al., 2020). The paper’s central finding is that high average accuracy can coexist with severe failure on particular bilingual tuples when accented speech is present.

Language conditioning is equally important in low-resource bilingual ASR. A cross-linguistic evaluation over nine language pairs fine-tunes XLS-R 1B on bilingual data by prepending each target transcription with a language identification token; at inference, the model jointly predicts both the language and the transcription from speech alone (Amooie et al., 16 Jun 2026). The analysis defines

$$\hat{\ell}(s) = \argmax_{\ell \in \mathcal{L}} p(\ell \mid s), \qquad p(\ell \mid s) = \frac{p(s \mid \ell)}{\sum_{\ell' \in \mathcal{L}} p(s \mid \ell')}.$$0

and reports an overall Pearson’s $$\hat{\ell}(s) = \argmax_{\ell \in \mathcal{L}} p(\ell \mid s), \qquad p(\ell \mid s) = \frac{p(s \mid \ell)}{\sum_{\ell' \in \mathcal{L}} p(s \mid \ell')}.$$1 with $$\hat{\ell}(s) = \argmax_{\ell \in \mathcal{L}} p(\ell \mid s), \qquad p(\ell \mid s) = \frac{p(s \mid \ell)}{\sum_{\ell' \in \mathcal{L}} p(s \mid \ell')}.$$2 between LID accuracy and ASR improvement across repeated experiments (Amooie et al., 16 Jun 2026). When LID accuracy is high, bilingual fine-tuning can help low-resource ASR; when LID accuracy is low, providing the true LID token at inference improves WER. This directly links BLID reliability to downstream recognition quality rather than treating language labels as a separate auxiliary output.

These results establish a recurring pattern. In both multilingual dictation and low-resource bilingual ASR, BLID accuracy is not merely a front-end diagnostic metric: it changes the effective search space of the downstream recognizer or locale selector. A plausible implication is that BLID systems should be evaluated jointly with the tasks they gate, especially when language confusions are asymmetric across accents, locales, or related language pairs.

5. Evaluation regimes, benchmarks, and error analysis

BLID evaluation is highly sensitive to the benchmark design. DIVERS-BENCH evaluates state-of-the-art LID systems across speech transcripts, web text, social media, children’s stories, and code-switched text, and introduces DIVERS-CS, a natural code-switching resource spanning 10 language pairs: English–German, English–Hindi, Chinese–English, English–Tamil, English–Malayalam, Egyptian–English, Basque–Spanish, Arabic–English, English–Indonesian, and English–Turkish (Ojo et al., 22 Sep 2025). For code-switching, it uses Full Match (both languages correct) and Partial Match (at least one language correct). The reported Full Match score is near zero across evaluated models, while Partial Match is around 48–50%, indicating that current systems often detect only one of the two languages and are biased toward the dominant or higher-resource language, often English (Ojo et al., 22 Sep 2025). The paper explicitly attributes part of this failure to the softmax limitation of monolingual-centric classifiers.

Short-string evaluation reveals a different but related weakness: confusion among related languages. A reproduction of Apple’s two-layer character-level bi-LSTM for 20 Latin-alphabet languages trims inputs to 50 characters for training and evaluates on 10-character strings, using a shared 150-dimensional character embedding space and weighted cross-entropy with AdamW (Toftrup et al., 2021). On Universal Dependencies short strings, the reproduced model reports weighted $$\hat{\ell}(s) = \argmax_{\ell \in \mathcal{L}} p(\ell \mid s), \qquad p(\ell \mid s) = \frac{p(s \mid \ell)}{\sum_{\ell' \in \mathcal{L}} p(s \mid \ell')}.$$3, macro-$$\hat{\ell}(s) = \argmax_{\ell \in \mathcal{L}} p(\ell \mid s), \qquad p(\ell \mid s) = \frac{p(s \mid \ell)}{\sum_{\ell' \in \mathcal{L}} p(s \mid \ell')}.$$4, $$\hat{\ell}(s) = \argmax_{\ell \in \mathcal{L}} p(\ell \mid s), \qquad p(\ell \mid s) = \frac{p(s \mid \ell)}{\sum_{\ell' \in \mathcal{L}} p(s \mid \ell')}.$$5, $$\hat{\ell}(s) = \argmax_{\ell \in \mathcal{L}} p(\ell \mid s), \qquad p(\ell \mid s) = \frac{p(s \mid \ell)}{\sum_{\ell' \in \mathcal{L}} p(s \mid \ell')}.$$6, and $$\hat{\ell}(s) = \argmax_{\ell \in \mathcal{L}} p(\ell \mid s), \qquad p(\ell \mid s) = \frac{p(s \mid \ell)}{\sum_{\ell' \in \mathcal{L}} p(s \mid \ell')}.$$7 (Toftrup et al., 2021). Its error analysis shows that mistakes are concentrated in Romance, West Germanic, and North Germanic confusion clusters. For BLID, especially with related languages, this means that top-$$\hat{\ell}(s) = \argmax_{\ell \in \mathcal{L}} p(\ell \mid s), \qquad p(\ell \mid s) = \frac{p(s \mid \ell)}{\sum_{\ell' \in \mathcal{L}} p(s \mid \ell')}.$$8 behavior can be much more informative than top-1 accuracy.

Benchmark construction itself has become an active research topic. MCS-350 contains 52K children’s stories in 350+ languages and explicitly annotates multilingual and bilingual documents, including stories with co-occurring languages labeled as $$\hat{\ell}(s) = \argmax_{\ell \in \mathcal{L}} p(\ell \mid s), \qquad p(\ell \mid s) = \frac{p(s \mid \ell)}{\sum_{\ell' \in \mathcal{L}} p(s \mid \ell')}.$$9 (Agarwal et al., 2023). It is used both as a benchmark for short-text LID and as a source for extracting code-mixed and parallel sentences. Because the dataset emphasizes underrepresented Indian and African languages, it broadens BLID evaluation beyond the standard high-resource pairs that dominate earlier literature.

The common theme across these benchmarks is that BLID is often masked by monolingual summary metrics. High $p(s \mid \ell)$0 on curated single-language data does not imply reliable bilingual set prediction, reliable span detection, or robust performance under domain shift. This is not an inferential leap but a repeated empirical outcome across short strings, children’s stories, social media, and code-switched sentences.

6. Persistent difficulties and research directions

A recurring misconception in the literature is that language identification is solved. One paper states that it is sometimes even considered solved, but that current systems cannot accurately identify most of the world’s 7000 languages because of lack of data and computational challenges (Agarwal et al., 2023). Another shows that high average accuracy for multilingual speakers can conceal sharp failures for accented or confusable language tuples (Titus et al., 2020). DIVERS-BENCH further shows that models that achieve high $p(s \mid \ell)$1 on clean monolingual benchmarks can collapse on bilingual or code-switched inputs, with Full Match near zero (Ojo et al., 22 Sep 2025). Taken together, these results make clear that BLID remains open wherever language similarity, accent variation, short context, code-switching, or low-resource coverage are central.

The literature points toward several convergent design principles. One is to replace monolithic monolingual classifiers with models that can express multiple languages or boundary structure: character-level span taggers, word-level sequence models with CRF decoding, multilabel or span-based evaluation, and frame-synchronous speech predictors all move in this direction (Kocmi et al., 2017, Mandal et al., 2018, Zhang et al., 2022). Another is to encode bilingual ambiguity explicitly rather than forcing hard labels, as in the wal-gof class for Wolaita/Gofa or the confusion-resolution units of LIMIT (Yigezu et al., 5 Sep 2025, Agarwal et al., 2023). A third is to exploit broader context—speaker turns, installed locales, selected locale, recent toggles, or external LID tokens—when the acoustic or lexical signal alone is insufficient (Valente et al., 2024, Titus et al., 2020, Amooie et al., 16 Jun 2026).

A further research direction is data efficiency. Cross-lingual self-supervised speech pre-training, tokenizer-based probabilistic BLID, and lightweight confusion-specific classifiers all report strong behavior under limited supervision (Tjandra et al., 2021, Meister et al., 19 Feb 2026, Agarwal et al., 2023). This suggests that future BLID systems may depend less on large bilingual annotated corpora and more on transferable multilingual pre-training plus targeted adaptation for confusion pairs.

BLID therefore occupies a distinctive position in multilingual NLP and speech processing. It is simultaneously a classification problem, a segmentation problem, a calibration problem, and, in deployed systems, a routing problem whose errors propagate into ASR, translation, moderation, and user-facing language selection. The strongest current results are obtained not by assuming a single universal solution, but by matching model structure and evaluation regime to the bilingual phenomenon at hand: short-string ambiguity, code-switched spans, accented multilingual speakers, turn-level language alternation, or low-resource pairs with genuine lexical overlap.