Spontaneous speech in multilingual communities such as India frequently combines code-switching (CS) and disfluencies, yet existing Bengali–English and Hindi–English speech corpora largely consist of fluent or scripted utterances. This limits their suitability for developing and evaluating automatic speech recognition (ASR) systems intended for real conversational settings, particularly in micro-resource scenarios. We introduce BEHE-CMDisfl, a synthetic speech corpus that explicitly integrates disfluency phenomena within Bengali–English and Hindi–English code-mixed (CM) utterances. The textual content was generated using prompting strategies with large language models (LLMs) to encourage controlled switching and varied disfluency patterns, including filled pauses, repetitions, and restarts. The utterances were subsequently synthesized using Indic Parler text-to-speech (TTS) system. To demonstrate usability, we establish a reproducible GMM–HMM baseline for Bengali–English ASR using Kaldi on a 1.3-hour subset of the corpus. In our experiments, improvements were mainly observed after ensuring consistency in the pronunciation lexicon and applying phonetic normalization, with the best setup reaching a word error rate (WER) of 37.74%. A closer look at the decoded transcripts suggests that filled pauses and repetitions are not automatically collapsed, but appear in the output, indicating that the disfluency cues present in the synthetic speech are captured during recognition.

Spontaneous speech in multilingual communities often involves rapid code-mixing (CM) and natural disfluencies, yet such patterns are rarely reflected in available training data for under-resourced languages. This gap limits the development of robust automatic speech recognition (ASR) systems. To address this, we introduce BEHE-CMDisfl, a fully synthetic Bengali–English and Hindi–English disfluent code-mixed speech corpus generated through a controlled Large Language Model (LLM) and Text-to-Speech (TTS) pipeline. The dataset explicitly incorporates conversational phenomena such as filled pauses, repetitions, and restarts. We evaluate its usefulness under two ASR settings. In a micro-resource scenario (∼1.3 hours), a GMM-HMM Kaldi baseline achieved a 37.74% Word Error Rate (WER) after phonetic normalization to reduce transliteration inconsistencies, and successfully retained disfluency markers in decoding. We also examined adaptation of a modern foundation model. In zero-shot testing, openai/whisper-small failed on the code-mixed speech due to severe hallucinations and looping behavior. After applying parameter-efficient fine-tuning (LoRA) for 1,000 steps, the model stabilized, reduced insertion errors, captured rapid language switching more reliably, and achieved a WER of 21.37%. These findings show that synthetic data combined with efficient fine-tuning offers a practical path for ASR development in complex low-resource disfluent CM settings.