Panhapich/khmer-english-codeswitch-tts

Khmer–English Code-Switch Synthetic Speech

8,177 utterances / 15.9 hours of synthetic Khmer–English code-switched speech, generated with VoxCPM2 from code-switch text manufactured by confirmed lexical substitution over a Khmer–English parallel corpus.

⚠️ This is synthetic speech, not recordings of people. Every utterance was produced by a TTS model, and the code-switch sentences were manufactured by word substitution — they are not transcripts of anything a person said. It is intended as an augmentation set to be mixed with real speech, following the TTS-augmentation literature for low-resource and code-switch ASR. See Limitations before training on it.

Why this exists

Public Khmer speech data is scarce — essentially OpenSLR 42 (4 h, male speakers only) and the Khmer ASR Cultural Dataset (106 h). Public code-switch speech corpora are Mandarin–English (SEAME, ASRU 2019, CS-Dialogue) or Hinglish (HiACC). At the time of construction, a search turned up no public Khmer–English code-switch speech corpus at all, while Khmer–English parallel text is abundant. Converting the cheap asset (text) into the expensive one (speech) is exactly what the TTS-augmentation literature is for.

How it was built

1. Source pairs. 4.2M Khmer–English sentence pairs, filtered to 1.36M by deduplication, script-ratio checks, length bounds, and rejection of spec-sheet/product-listing fragments.
2. Segmentation. The Khmer side is segmented with khmercut.
3. Alignment. eflomal aligns both directions, symmetrised with grow-diag-final-and, yielding a KM→EN lexicon filtered to high-confidence content words (bidirectional p ≥ 0.20, count ≥ 10, English side non-stopword). Proper nouns are removed via sentence-medial capitalisation rate, except an explicit allowlist of terms Khmer speakers genuinely code-switch (country and nationality names, internet, facebook, …) — this is what keeps romanised Khmer name fragments like penh or sen out of the data.
4. Substitution — confirmed twice. A Khmer token is swapped only if it maps to an English word in the lexicon and that English word literally appears in that sentence's own human translation. The lexicon proposes; the human translator confirms.
5. Switch density. 1–3 non-adjacent, never sentence-initial swaps, targeting a Code-Mixing Index band of 5–25 % (Gambäck & Das, LREC 2016).
6. Synthesis. VoxCPM2 (2B, 48 kHz native, bf16), 10 CFM steps, cfg 2.0, one of 20 Voice Design speaker descriptors assigned deterministically per utterance. Audio is resampled to 24 kHz FLAC.
7. Filtering. Signal-level QA: duration-vs-text-length outliers (catching truncation and runaway generation), silence, clipping, RMS floor, NaN.

Source sentences that already contained English were excluded, so every English token in cs_text is a controlled insertion whose position is recorded in swaps.

Statistics

Utterances	8,177
Duration	15.9 hours
Mean / median utterance	7.0s / 6.4s
Mean / median CMI	9.8 / 8.7
Mean swaps per utterance	1.45
Distinct synthetic voices	20
Distinct English words inserted	1,141
Total English insertions	11,881
Sample rate	24 kHz FLAC

Most frequently inserted English words: cambodia, new, national, government, water, international, high, khmer, law, ensure, health, quality, public, know, best

Fields

field	type	description
audio	Audio(24 kHz)	synthesised waveform
cs_text	string	the code-switched sentence that was synthesised — the transcript
km_orig	string	original monolingual Khmer sentence
en_orig	string	its human English translation
swaps	list[string]	each substitution as khmer_token->english_word
n_swaps	int	number of substitutions
cmi	float	Code-Mixing Index of the utterance
voice_desc	string	Voice Design descriptor used
duration_s	float	audio duration
seed	int	RNG seed (generation is reproducible)
pair_id	string	stable id derived from the source sentence

from datasets import load_dataset
ds = load_dataset("Panhapich/khmer-english-codeswitch-tts", split="train")
print(ds[0]["cs_text"], ds[0]["swaps"], ds[0]["cmi"])

Limitations

Read these before using the data.

• Word substitution is not attested code-switching. Real code-switching follows syntactic constraints and sociolinguistic habit. Lexical substitution produces plausible switch points, not attested ones. Occasional artifacts remain where a Khmer multi-word construction is split (e.g. swapping the "good" of "good-most" for best, stranding the modifier).
• Synthetic speech is less diverse than real speech, and naively mixing it with real data can underperform (Ogun et al., 2025). A synthetic-to-real representation gap persists even when audio sounds good (Su et al., 2024; Quintas et al., 2024).
• Speaker diversity is bounded by 20 synthetic Voice Design descriptors. These are not 20 real speakers, and they do not model Khmer regional accents.
• Khmer-side quality is not ASR-verified. Every available Khmer ASR is weak and Khmer-only, so it would mangle exactly the English insertions that matter; a round-trip filter would delete good data as readily as bad. Only assumption-free signal-level QA was applied. Sample the audio yourself before trusting it at scale.
• Domain follows the source corpus — news, government, commerce, general web text. It is written-register text read aloud, not spontaneous conversation, which is where real code-switching lives. Expect a domain gap against conversational CS benchmarks.
• The base TTS model's Khmer is not independently evaluated. Khmer is 1 of 30 languages in VoxCPM2, whose published benchmarks are dominated by Chinese and English.

Recommended use

Mix with real speech rather than training on it alone; every study in the bibliography that succeeded did so by augmenting real data. Staged training or parameter regularisation is advisable to avoid catastrophic forgetting (Fazel et al., 2021).

Provenance and licensing

• Source text: a combined English–Khmer parallel corpus (CC-BY-4.0), merging publicly available EN–KM datasets.
• TTS model: openbmb/VoxCPM2, Apache-2.0.
• This dataset: CC-BY-4.0, inherited from the source text.

The repo also carries research/REFERENCES.md (annotated bibliography), research/NOTES.md (the argument and its weaknesses), benchmarks/REPORT.md (measured generation throughput) and qa_report.md (filtering results).

Key references

• Rossenbach et al. (2020), Generating Synthetic Audio Data for Attention-Based Speech Recognition Systems — arXiv:1912.09257
• Sharma et al. (2020), Improving Low Resource Code-switched ASR using Augmented Code-switched TTS — arXiv:2010.05549
• Ibaraki & Chiang (2025), Frustratingly Easy Data Augmentation for Low-Resource ASR — arXiv:2509.15373
• Yeo et al. (2026), Improving Code-Switching ASR with Code-Mixing Guided Synthetic Speech — arXiv:2606.19381
• Gambäck & Das (2016), Comparing the Level of Code-Switching in Corpora — ACL L16-1292