Dulu Malayalam ASR

Dulu Malayalam ASR

Turns spoken Malayalam into text on a CPU or a CUDA GPU, as an early research release from Serendipity Labs. For people who want draft transcripts of short, clear recordings.

Made by serendipitylabs on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Dulu turns spoken Malayalam into text. Serendipity Labs published it as an early research release, and its base is Google's EmbeddingGemma 2. All the weights ship with it, about 1.3 GB, so you do not fetch the base model separately. Treat each transcript as a draft that needs a human check, especially for names and numbers.

Who it suits: It suits people who want a first-draft text of Malayalam speech, and researchers who work on speech recognition for this language.

Industry | Who uses it | How they use it | Example
Media | Editors and journalists | Draft a transcript of a short, clear clip, then review it | Getting a first text of a spoken interview excerpt
Research | Speech researchers | Reproduce the card's Kathbath and FLEURS scores | Comparing another Malayalam model against the reported error rates
Education | Teachers and learners | Turn short spoken lessons into text to check by hand | Reading back a recorded sentence in written form

If you need live captions, punctuation, speaker labels or timestamps, you can skip it, because it provides none of them.

Where to use it:

• Windows, macOS or Linux with Python 3.12 or 3.13
• A CPU, or a CUDA GPU that supports BF16
• Your own Python code, importing DuluASR from the downloaded folder

Use case | Who | What they get back
Transcribe one recording | Anyone with Malayalam audio | The recognized text for a wav file
Call it from code | Developers | A transcript printed from model.transcribe
Check its accuracy | Researchers | Word and character error rates on two evaluation sets

How to set it up:

• Python 3.12 or 3.13
• The huggingface_hub package at version 2.2.0
• Disk space for about 1.3 GB of weights, plus extra memory while it runs
• A short recording of clear speech, which the tool changes to 16 kHz mono by itself

1. Install the pinned huggingface_hub package.
2. Download the model into a folder named dulu-malayalam-asr.
3. Move into that folder.
4. Install its requirements.
5. Run the script on your recording.

python -m pip install huggingface_hub==2.2.0
hf download serendipitylabs/dulu-malayalam-asr --local-dir dulu-malayalam-asr
cd dulu-malayalam-asr
python -m pip install -r requirements.txt
python dulu_asr.py recording.wav --device cpu

How to use it: Run the script with your recording and a device. Use --device cpu for the processor, or --device cuda on a CUDA GPU that supports BF16. From Python, work inside the downloaded folder.

import torch
from dulu_asr import DuluASR

torch.set_num_threads(4)
model = DuluASR.from_pretrained(".", device="cpu")
print(model.transcribe("recording.wav"))

Tool | What it does
dulu_asr.py | Transcribes a recording from the command line
--device | Chooses cpu or cuda for the run
DuluASR.from_pretrained | Loads the model from the downloaded folder
model.transcribe | Returns the text for a recording

Example: You run the script on a short, clear Malayalam recording saved as recording.wav, with the cpu device. The Malayalam text comes back, and you read it as a draft and fix names, numbers or unfamiliar words.

For newcomers and experts: Speech recognition means a program listens to a recording and writes down what was said. A good first try is a short recording of one person speaking clearly, run with the cpu device.

Under the hood, the training set was Kathbath, 134.1 hours of Malayalam speech. The checkpoint was chosen on the Kathbath validation split, and FLEURS served as the final test. The release is epoch 4 of a 12-epoch run. Character error rates are 7.15 percent on Kathbath validation and 11.15 percent on FLEURS. The scores come from one run, with text normalized first, greedy decoding and no language model.

Recording is converted to mono 16 kHz
The model reads the audio
Malayalam text is produced
You review the draft

Good to know:

Title: Word error rate, lower is better
Kathbath validation: 41.07%
FLEURS Malayalam test: 44.13%

• Kathbath validation includes speakers who also appear in training, which can flatter the first score.
• This is a research release. Conversations, dialects, Malayalam and English mixed together, noisy audio and long recordings are not thoroughly tested.
• CPU and GPU outputs can differ slightly.
• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the audio files you give it, run on your own machine, and the weights fetched from the Hugging Face hub
• Not for: production transcription, punctuation, speaker identification, timestamps or live audio streams

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Library: pytorch
• Downloads in the last 30 days: 71
• Likes: 2
• License: apache-2.0
• Languages: ml
• Based on: google/embeddinggemma-2
• Tags: pytorch, malayalam, speech, gemma, research, automatic-speech-recognition, model-index

Links

• Model: https://huggingface.co/serendipitylabs/dulu-malayalam-asr
• Files: https://huggingface.co/serendipitylabs/dulu-malayalam-asr/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download serendipitylabs/dulu-malayalam-asr

Then follow the usage in its model card. Nothing in this kit runs by itself.