Panhapich/khmer-english-codeswitch-tts-llm

Khmer–English Code-Switch Synthetic Speech (LLM-authored)

19,825 utterances / 21.7 hours of synthetic Khmer–English code-switched speech at 16 kHz, generated with VoxCPM2 from code-switch sentences written by an LLM and validated programmatically.

⚠️ This is synthetic speech, not recordings of people. Every utterance was produced by a TTS model, and every sentence was written by a language model — 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. Converting a cheap asset into the expensive one is exactly what the TTS-augmentation literature is for.

This corpus is the successor to a substitution-built version of the same idea.

How it was built

1. Text generation. An LLM (deepseek-v4-flash:preview via the Ollama API) is asked for natural Khmer–English code-switched sentences — the sentence and nothing else. The prompt is assembled from four crossed axes sampled per request: 50 domains × 18 speaking situations × 6 registers × 5 sentence forms. A single fixed prompt run thousands of times returns thousands of variations on the same handful of sentences; the axes are what stop the corpus collapsing onto one topic and one register.

2. Validation. Nothing the model returns is trusted. A model asked to code-switch will happily return monolingual Khmer, a fully English sentence, or — worst, because it looks like success — Khmer romanised into Latin letters. Every candidate must clear, in order:

   • both scripts present (a monolingual answer is not a code switch);
   • every Latin token found in an English vocabulary built from a 1.36M-pair Khmer–English parallel corpus (110k word forms), which is what catches romanised Khmer;
   • Code-Mixing Index inside a 5–25 % band (Gambäck & Das, LREC 2016), computed with khmercut segmentation on the Khmer side and Latin word runs on the English side;
   • length bounds; no digits, URLs, markup, invisible characters, or degenerate repetition, all of which a TTS front-end reads unpredictably;
   • not a duplicate, normalised, of anything already accepted.

   Sentences are respaced at every Khmer↔Latin script transition: Khmer is written unspaced and models routinely butt the insertion straight against it, which the TTS front-end mis-segments.

   73.1% of 27,359 candidates were accepted. Rejection reasons are counted and reported rather than silently discarded — see Rejections.

3. Synthesis. VoxCPM2 (2B, 48 kHz native, bf16), 10 CFM steps, cfg 2.0, one of 20 Voice Design speaker descriptors assigned deterministically per utterance from a hash of its id, so voice and RNG seed are reproducible. Audio is resampled to 16 kHz FLAC — the rate ASR training consumes — through a filtered torchaudio resampler.

4. Filtering. Signal-level QA: duration-vs-text-length outliers (catching truncation and runaway generation), silence, clipping, RMS floor, NaN. 19,825 of 20,000 synthesised utterances survived — 174 duration outliers and 1 too-quiet clip were dropped. ASR round-trip filtering (Tier 2) was not applied; see Limitations.

Statistics

Utterances	19,825
Duration	21.7 hours
Mean / median utterance	3.9s / 3.8s
Mean / median CMI	15.5 / 15.4
Mean English insertions per utterance	1.36
Distinct synthetic voices	20
Distinct English words/phrases inserted	6,817
Total English insertions	26,996
Sample rate	16 kHz FLAC, 16-bit

Most frequently inserted English: deadline, meeting, schedule, budget, discount, order, project, contract, ticket, homework, scholarship, prescription, team, passport, delivery

Rejections

What the validator threw away, and why. A model that cannot do this task shows up here rather than in the corpus:

reason	count	share of candidates
cmi_out_of_band	4,662	17.0%
too_few_khmer_tokens	1,539	5.6%
romanised_or_unknown_english	377	1.4%
no_english	290	1.1%
digits	242	0.9%
english_initial	152	0.6%
truncated_response	43	0.2%
length	25	0.1%
request_failed	14	0.1%
junk_chars	10	0.0%
duplicate	3	0.0%

cmi_out_of_band dominating is expected and healthy — it is a density knob, not a defect, and the model simply overshoots or undershoots the 5–25 % target more often than it fails at anything else.

Fields

field	type	description
audio	Audio(16 kHz)	synthesised waveform
cs_text	string	the code-switched sentence that was synthesised — the transcript
swaps	list[string]	each English insertion as ?->english; the ? is a placeholder, since the model wrote the sentence rather than substituting into a Khmer original
n_swaps	int	number of English insertions
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, a hash of the sentence
km_orig, en_orig	string	empty in this corpus. Present for schema compatibility with the substitution corpus, where they hold the Khmer original and its human translation

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

Relationship to the substitution corpus

An earlier corpus built the same way downstream, but manufactured its text by confirmed lexical substitution: a Khmer word was replaced by its English counterpart only when a corpus-derived alignment lexicon proposed it and that English word literally appeared in that sentence's own human translation. That method is safe and fully traceable — every swap has a human translator behind it — but it is word-for-word by construction, so it cannot produce the multi-word English phrases that dominate real Cambodian code-switching, and it inherits the source corpus, which is heavy with scraped product listings and scripture.

The trade is explicit: the substitution corpus buys per-sentence human confirmation at the cost of naturalness; this one buys naturalness at the cost of that confirmation. Neither is attested speech.

Limitations

Read these before using the data.

• The sentences are LLM output, not attested code-switching. They are what a model believes bilingual Cambodians say. The validation checks structural properties — script mix, switch density, vocabulary — not sociolinguistic authenticity. No native speaker has reviewed this corpus at scale. That review is the single most valuable thing anyone could add to it.
• The romanisation check is a heuristic. Its vocabulary is the English side of a scraped parallel corpus, which itself contains some romanised Khmer, so the check catches the bulk case rather than every case.
• 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. Because the model was asked for the sentence alone, there is also no monolingual Khmer reference text to score a round trip against. Only assumption-free signal-level QA and English-insertion recovery were applied. Sample the audio yourself before trusting it at scale.
• Register is spoken-style but not spontaneous. The prompt asks for conversational Khmer and gets it, but this is still written text read aloud by a TTS model — no disfluencies, no overlapping speech, no real prosodic spontaneity.
• Topic distribution is the prompt grid's, not the world's. 50 domains chosen by hand is a broader and cleaner spread than the source corpus offered, but it is still a hand-chosen spread.
• 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

• Text: generated by deepseek-v4-flash:preview via the Ollama API. The English vocabulary used for validation derives from a combined English–Khmer parallel corpus (CC-BY-4.0).
• TTS model: openbmb/VoxCPM2, Apache-2.0.
• This dataset: CC-BY-4.0.

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