ASR models

Four speech recognition models tuned for different domains. Pass the name in the model field. You can't mix models within a single request, but you can switch between requests freely.

Side-by-side

zero-indic

The default for Indian languages. Native handling of Hindi, Tamil, Telugu, Kannada, Marathi, Bengali, Gujarati, Punjabi, Malayalam, Odia, Urdu, Assamese, Maithili, and more.

When to use

• Single-language Indic audio where you know (or can detect) the language
• Call-centre audio for Indian customers in their native tongue
• Meetings, interviews, dictation, media transcription

Example

curl -X POST https://asr.shunyalabs.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer $SHUNYALABS_API_KEY" \
  -F "file=@call.wav" \
  -F "model=zero-indic" \
  -F "language_code=hi"

zero-universal

99-language Whisper-class foundation model (docs say 204; the published leaderboard number varies by evaluation set). Strongest on English and European languages. Use when the input could be anything.

When to use

• English or non-Indian-language content
• Audio where you don't know the language in advance, set language_code=auto
• International meetings with mixed English and other languages

Benchmarks

Composite WER on the HuggingFace Open ASR leaderboard: 3.10%, claimed 48% fewer errors than the next-best model.

Example

curl -X POST https://asr.shunyalabs.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer $SHUNYALABS_API_KEY" \
  -F "file=@podcast.wav" \
  -F "model=zero-universal"

zero-med

Clinical-grade. Trained to recognise drug names, procedures, anatomy, dosages, and diagnostic terms. Auto-applies medical terminology correction via MedGemma (no extra flag needed).

When to use

• Doctor-patient consultations
• OT / surgical notes dictation
• Case note transcription for EHR ingest

Compliance

Example

curl -X POST https://asr.shunyalabs.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer $SHUNYALABS_API_KEY" \
  -F "file=@consult.wav" \
  -F "model=zero-med" \
  -F "language_code=en"

Output shape

Same response schema as other models; the value-add is in the text itself, drug names in correct spelling, procedure names in standard terminology. Pair with enable_keyterm_normalization for further normalization against your own glossary.

zero-codeswitch

Tuned for speech that mixes two languages within a single utterance, Hinglish (mujhe yeh passbook update karna hai), Tanglish, Banglish, etc. Automatically restores English words to Latin script so the final transcript reads cleanly.

When to use

• Urban Indian call-centre conversations (customers routinely switch mid-sentence)
• Social-media, podcast, or creator content with code-mixed speech
• Any conversation where separating languages would force unnatural phrasing

Example

curl -X POST https://asr.shunyalabs.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer $SHUNYALABS_API_KEY" \
  -F "file=@hinglish.wav" \
  -F "model=zero-codeswitch"

What the output looks like

{
  "text": "मुझे अपना EMI details check करना है",
  "detected_language": "Hinglish",
  "segments": [
    { "start": 0.3, "end": 3.1, "text": "मुझे अपना EMI details check करना है" }
  ]
}

Picking the right model, decision flow