Speech ➜ Pure Roman Hinglish · Market Survey

Yes. Hinglish ASR that outputs pure Roman script — it exists.

Feed it Hindi–English code-switched speech, get back Latin-script Hinglish only — zero Devanagari. Below is every model found on the market (as of September 2026): open-weight checkpoints you can download today, plus closed-weight APIs, with parameter counts, FP32 sizes, quantization sizes and published Hinglish WER.

input ▸ "bas din mein kitni baar chalti hai?" output ▸ vah bas din mein kitni baar chalti hai? · 100% Roman

"Mehnat to poora karte hain." "Haan vahi ek aapko bataaya na." "Aap pandrah log hain." "Nahin, just thank you, thank you."

▲ real transcriptions from Oriserve Whisper-Hindi2Hinglish models — Roman script, no Devanagari mix

Short answer

YES ✓

Native models exist — both fine-tuned and code-switching-native APIs.

Open weight

– repos

Downloadable checkpoints + ready quantizations (GGML · CT2 · MLX · CoreML · ONNX).

Closed weight

– APIs

Sarvam Saaras (translit / codemix) & Deepgram Nova-3 / Flux multilingual.

Best published Hinglish WER

15.85%

moorlee/qwen3-asr-0.6b-hinglish on HiACC (conversational Hinglish).

🎙️ Open-weight models — Roman-script Hinglish

Weights are fully downloadable. Every model here natively emits Roman/Latin script for Hindi–English code-switched speech (no post-hoc transliteration step). trained = original fine-tune / training run.

Model	Type	Base / architecture	Params	FP32 size	Quantizations available	Hinglish WER*	License

* WERs are self-reported by each author on different benchmarks and are NOT directly comparable across rows. Benchmarks: FLEURS = google/fleurs hi_in test · CV = Common Voice 20 hi · IV = IndicVoices hi · HiACC = HiACC conversational Hinglish · MUCS = OpenSLR-104 / MUCS-2021 Hinglish. Sizes marked ≈ are estimates from bytes/parameter (FP32 = 4 B, FP16 = 2 B, INT8 ≈ 1.06 B, Q5_1 ≈ 0.71 B, Q4 ≈ 0.57 B).

📦 Ready-made conversions & quantizations

Drop-in runtime formats of the trained models above — for whisper.cpp, faster-whisper / CTranslate2, Apple MLX & WhisperKit, sherpa-onnx, and ONNX deployments.

🔒 Closed-weight models — Roman-script Hinglish

APIs only — weights are not public, so params / FP32 / quant sizes are undisclosed. All output Latin script only (no Devanagari) for Hindi–English code-switched audio.

Model	Vendor	How to get Roman output	Params / FP32 / Quants	Hinglish WER*	Access

* Sarvam WER figures are from the official VISTAAR benchmark (transcribe mode); WER for the translit / codemix output modes is not separately published. Deepgram does not publish Hindi/Hinglish WER.

🚫 Checked & excluded — they fail the pure-Roman test

Popular ASR systems that handle Hindi (some even code-switching) but output Devanagari for Hindi speech, or cannot guarantee Latin-only Hinglish — excluded per the criteria.