EkaCare Parrotlet-a 2.5 Pro
Parrotlet-a 2.5 Pro is a purpose-built automatic speech recognition (ASR) model for medical speech in Indian healthcare settings. It transcribes Indian English, Hindi, Marathi, Kannada and Telugu, including the heavily code-mixed speech typical of real consultations (English drug names and clinical terms embedded in Indic speech).
The model combines a Whisper large-v3 encoder with a MedGemma 4B decoder through a lean projector layer, and was further tuned with GRPO on medical conversation data. Weights are stored in bfloat16.
Benchmark results
Re-benchmarked against nine production ASR systems on medical evaluation sets in five Indian languages, plus AI4Bharat's IndicVoices as an out-of-domain check. Every medical sample carries a SQUIM objective STOI-based audio-quality rating (binned into Excellent / Good / Bad / Poor) and medical-entity annotations.
Metric — Semantic WER: word error rate after medical-unit, number, transliteration and orthography normalization, so a model is not penalised for spelling a drug name in a different script. Lower is better.
Semantic WER — medical conversations
| Model | Indian English | Hindi | Marathi | Kannada | Telugu |
|---|---|---|---|---|---|
| Parrotlet-a 2.5 Pro (Eka.care) | 9.15 | 18.16 | 31.61 | 29.24 | 24.77 |
| Gemini 3.1 Pro (Google) | 12.09 | 19.08 | 39.26 | 36.38 | 31.80 |
| Gemini 3.6 Flash (Google) | 12.50 | 21.96 | 44.44 | 43.43 | 37.81 |
| GPT Transcribe (OpenAI) | 13.00 | 26.71 | 51.98 | 58.92 | 42.56 |
| Saaras v3 (Sarvam) | 14.59 | 21.64 | 40.75 | 33.33 | 29.79 |
| Scribe V2 (ElevenLabs) | 16.23 | 21.91 | 38.14 | 39.61 | 33.40 |
| Whisper Large v3 (OpenAI) | 17.46 | 52.26 | 101.42 | 99.59 | 96.22 |
| Prisma v2.5 (Gnani) | 26.12 | 35.90 | 52.08 | 51.99 | 53.82 |
| IndicConformer 600M (AI4Bharat) | — | 37.49 | 56.02 | 58.37 | 42.09 |
Best in every language on the medical conversation sets.
Medical entity accuracy
Share of medical keyword tokens recovered — drug names, doses, conditions, procedures. This is the number that decides whether a transcript is clinically usable. Higher is better.
| Model | Indian English | Hindi | Marathi | Kannada | Telugu |
|---|---|---|---|---|---|
| Parrotlet-a 2.5 Pro (Eka.care) | 95.05 | 89.10 | 82.46 | 86.01 | 88.99 |
| Gemini 3.1 Pro (Google) | 94.94 | 88.08 | 57.43 | 70.38 | 71.55 |
| Gemini 3.6 Flash (Google) | 94.00 | 82.88 | 52.75 | 63.94 | 64.08 |
| GPT Transcribe (OpenAI) | 90.96 | 73.58 | 42.19 | 43.69 | 54.44 |
| Saaras v3 (Sarvam) | 87.23 | 76.71 | 46.11 | 62.87 | 63.18 |
| Scribe V2 (ElevenLabs) | 86.16 | 77.64 | 69.18 | 75.29 | 79.42 |
| Whisper Large v3 (OpenAI) | 83.81 | 47.42 | 17.37 | 8.50 | 19.95 |
| Prisma v2.5 (Gnani) | 72.26 | 50.93 | 32.31 | 42.37 | 40.63 |
| IndicConformer 600M (AI4Bharat) | — | 50.36 | 27.37 | 35.73 | 44.76 |
Semantic WER — IndicVoices (out-of-domain)
General-domain read and spontaneous speech, outside the medical domain the model is tuned for. Parrotlet-a 2.5 Pro leads on Kannada and sits within a point or two of the best models on Telugu and Marathi — specialisation has not cost general coverage.
| Model | Hindi | Marathi | Kannada | Telugu |
|---|---|---|---|---|
| Saaras v3 (Sarvam) | 10.54 | 13.16 | 25.53 | 20.57 |
| Prisma v2.5 (Gnani) | 11.65 | 14.20 | 25.16 | 19.20 |
| IndicConformer 600M (AI4Bharat) | 10.76 | 12.58 | 26.25 | 20.69 |
| Parrotlet-a 2.5 Pro (Eka.care) | 12.45 | 13.27 | 24.71 | 20.01 |
| Gemini 3.1 Pro (Google) | 12.51 | 18.51 | 33.72 | 24.51 |
| Scribe V2 (ElevenLabs) | 12.21 | 19.58 | 38.66 | 28.08 |
| Gemini 3.6 Flash (Google) | 15.99 | 24.04 | 41.72 | 30.69 |
| GPT Transcribe (OpenAI) | 14.24 | 23.27 | 47.12 | 28.96 |
| Whisper Large v3 (OpenAI) | 25.52 | 77.91 | 88.79 | 105.53 |
Audio-quality robustness
On the medical sets split by audio quality (Excellent → Poor, estimated from the audio itself using SQUIM objective STOI, binned at 0.5 / 0.65 / 0.85), Parrotlet-a 2.5 Pro keeps the flattest degradation profile and is the best model in nearly every quality bucket in every language — e.g. Indian English semantic WER of 6.49 on Excellent audio and 25.14 on Poor audio, versus 8.59 → 43.86 for the next-best model.
Installation Requirements
Python >= 3.10 and the following packages:
pip install "torch>=2.7.0" "transformers>=4.52.0,<5" librosa huggingface_hub
Loading the model from Hugging Face Hub
from transformers import AutoModel
import librosa
repo_name = "ekacare/parrotlet-a-2.5-pro"
model = AutoModel.from_pretrained(repo_name, trust_remote_code=True)
Load an audio file
audio_path = "path/to/your/audio.mp3"
audio, sample_rate = librosa.load(audio_path, sr=16000) # 16 kHz is required — sr=16000 resamples on load
Perform speech recognition
transcription = model.transcribe(audio, sample_rate)
print("Transcription:", transcription)
Notes
- Audio (wav, mp3) must be passed at 16 kHz — load with
librosa.load(..., sr=16000)as above. Other sample rates are not currently handled bytranscribe(). - The model handles short-form audio up to 30 seconds; longer clips are silently truncated — chunk longer recordings yourself before transcribing.
- Weights are bfloat16; the model runs on GPU (CUDA) or CPU.
License
This model combines three components:
- Audio encoder — derived from
openai/whisper-large-v3(Apache-2.0). - Projector layers — trained from scratch by Eka Care (Apache-2.0).
- Decoder — a fine-tune of
google/medgemma-4b-it, a model derivative under Google's Health AI Developer Foundations Terms of Use.
The combined weights are therefore distributed under the Health AI Developer Foundations Terms of Use — by downloading or using this model you agree to those terms, including their use restrictions and the requirement that further derivatives carry the same terms.
Citation
If you use this model, please cite:
@software{parrotlet_a_2_5_pro,
author = {{Eka Care}},
title = {Parrotlet-a 2.5 Pro: medical speech recognition for Indian languages},
year = {2026},
url = {https://huggingface.co/ekacare/parrotlet-a-2.5-pro}
}
- Downloads last month
- 2