whisper-hinglish-preview

Whisper Hinglish (Preview)

A Whisper-large-v3 model specialised for Hinglish (Hindi–English code-switched) speech, with strong pure-Hindi and English transcription. It is the best-performing open weights model in our internal evaluation across code-switch, Hindi, and English benchmarks.

Research preview. Numbers and weights may change. Evaluated on internal benchmarks; see disclaimers below.

---

Try it live

Available on Trelis Router (https://router.trelis.com/models):

• UI — log in and upload an audio clip to transcribe right in the browser.
• API — POST https://router.trelis.com/api/v1/transcribe for programmatic access (requires an API key).

Evaluation

Corpus WER (%, lower is better) under a script-safe indic-hindi normaliser (NFC + Indic normalisation, keeps Devanagari matras/nuktas, strips punctuation; not the Whisper default, which strips matras and inflates Devanagari WER). Compared against two leading commercial APIs: Sarvam (Saaras-v3) and ElevenLabs Scribe-v2.

🟠 Hinglish — code-switched (Hindi + English in one utterance, each in their native script)

🔵 Hindi (pure Devanagari)

⚪ English

Bold = best on that row.

---

How to use

Like any Whisper model, specify the language when you transcribe.

from transformers import WhisperProcessor, WhisperForConditionalGeneration
import soundfile as sf, torch

repo = "Trelis/whisper-hinglish-preview"
proc = WhisperProcessor.from_pretrained(repo)
model = WhisperForConditionalGeneration.from_pretrained(repo, torch_dtype=torch.bfloat16).to("cuda").eval()

audio, sr = sf.read("clip.wav")  # 16 kHz mono
feat = proc.feature_extractor(audio, sampling_rate=16000, return_tensors="pt").input_features.to("cuda", torch.bfloat16)

# Hindi audio → force <|hi|> ; English audio → force <|en|>
ids = proc.tokenizer.convert_tokens_to_ids
prompt = [ids("<|startoftranscript|>"), ids("<|hi|>"), ids("<|transcribe|>"), ids("<|notimestamps|>")]
out = model.generate(input_features=feat,
                     decoder_input_ids=torch.tensor([prompt]).to("cuda"),
                     max_new_tokens=440)
print(proc.tokenizer.decode(out[0], skip_special_tokens=True))

Code-switched audio. The model uses a dedicated <|mixedcode|> marker/token for utterances that mix Devanagari and Latin script. Insert it right after the language token, choosing the language token by the dominant script of the utterance:

mc = proc.tokenizer("<|mixedcode|>", add_special_tokens=False).input_ids
prompt = [ids("<|startoftranscript|>"), ids("<|hi|>"), *mc, ids("<|transcribe|>"), ids("<|notimestamps|>")]

---

Disclaimers

• Commercial-API WERs on pure Hindi benchmarks here are pessimistic. Sarvam and Scribe keep English loanwords in Latin script and numbers as digits, whereas our references render everything in Devanagari. A translit-blind WER then charges a substitution per loanword/number against them. The comparison is apples-to-apples on our Devanagari-reference protocol, not a claim about their raw quality.
• ᶜᵐ Sarvam evaluated in its code-mixed mode.
• Specify the language (<|hi|> / <|en|>) as shown above — standard Whisper usage — for the reported quality.

---

Attributions

• Architecture base: openai/whisper-large-v3.
• Starting checkpoint — Whisper-Vaani. Our Hindi/Hinglish training started from ARTPARK-IISc/whisper-large-v3-vaani-hindi, a Vaani-fine-tuned Whisper-large-v3 from the Vaani project (ARTPARK @ IISc). We gratefully credit the Whisper-Vaani model and the Vaani team.
• Evaluation benchmark: CoSHE-500 is derived from soketlabs/CoSHE-Eval (Soket Labs, CC-BY-NC-4.0).