Hima-ASR:
Open Indic Speech, and the Data Behind It
An open speech model for 11 Indian languages and English, small enough for a laptop, and unusually candid about its data: 34 open datasets and 132,130 scraped hours. Here is the model, every dataset grouped by how its text was made, and how you would train your own.
Updated: November 10, 2026 · Source: the Hima-ASR model card and its author · Read: 11 minutes
At a glance
- What it is: Hima-ASR, an open speech recognition model for 11 Indian languages and English, at 0.46B parameters. A small extra head (17.5M) answers plain-English questions about the same clip, so one listen gives you a transcript and decisions.
- Who built it: Prahaladha Reddy. The author's profile says MTS at Maya Research, previously an AI engineer at Kairos (acquired by OpenAI).
- How it does: on the card's own numbers, the lowest word error rate of four models in Telugu, Tamil, Kannada and Malayalam, against two 1.2B open models and Sarvam's API. Weaker in Hindi, Bengali and English.
- The interesting part: the data. The author says it was trained on 132,130 hours of scraped and relabelled audio, plus about 49,000 hours from 34 open datasets. Only 22,121 of those open hours were typed by people.
- Why it matters: in the author's words, the recipe is "mostly open and anyone with compute can go head to head or atleast very close to big labs". Below we list every dataset and sketch how you would train your own.
What Hima-ASR is
The model card's tagline is "speech to text, and speech to decisions, in one open model." It takes 16 kHz mono audio, 1 to 30 seconds per clip, and transcribes Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu and English. The card says a clip takes under half a second on a laptop GPU.
It is not trained from scratch. It is two existing open models joined by a small adapter, then trained on Indian speech:
12.5 frames per second
63,335 pieces kept
choice, yes/no or scale
Drawn from the architecture section of the model card. The encoder runs once per clip; the decoder and the decision head both read its output.
Two details are worth noting. The decoder keeps only the 63,335 tokenizer pieces it needs for Indian scripts and English, out of Gemma's 262,144, which is most of why the model is 0.46B rather than 0.59B. And because the decision head attends over audio frames, not the transcript, it can pick up tone, pauses and voice. Ask with no questions and it answers a ready-made set of 16 (tone, emotion, sentiment, intent, topic, urgency, language, speaker and more).
Using it
pip install "transformers>=5.13" torch torchaudio soundfile safetensors huggingface_hub
import torch
from transformers import AutoModel
model = AutoModel.from_pretrained("Prahaladha/hima-asr",
trust_remote_code=True, dtype=torch.float32)
print(model.transcribe("call.wav", language="hi"))
print(model.transcribe("call.wav", language="hi", style="mixed")) # English words in Latin letters
For decisions, the card's decide package exposes Hima("both"), whose run() takes a clip and a list of questions and returns the text plus one {option: probability} per question. It loads with trust_remote_code=True, so read the model code before running it in production.
Runtimes
There are ready-made ONNX and MLX versions that need no PyTorch. Speeds are from the card, measured on a MacBook Pro with an M4 Pro, in times realtime:
| Runs on | Version | Size | Speed |
|---|---|---|---|
| CPU | PyTorch | 920 MB | 16x |
| CPU | ONNX fp32 / int8 / q4 | 1.8 GB / 487 MB / 360 MB | 25x / 35x / 31x |
| Mac GPU | PyTorch | 920 MB | 21x |
| Mac GPU | MLX fp16 / 8-bit / 4-bit | 1.0 GB / 536 MB / 379 MB | 63x / 73x / 73x |
Benchmarks, as reported
The card scores four models on the same clips with the same scorer, language given, after standard normalisation. Word error rate in percent, lower is better. These are the author's numbers; we have not rerun them.
| Language (overall) | Hima-ASR 0.46B | Flex 1.2B | Core 1.2B | Sarvam Saaras v3 |
|---|---|---|---|---|
| All 12, average | 12.95 | 13.23 | 12.84 | 12.79 |
| Telugu | 13.4 | 15.0 | 14.6 | 15.5 |
| Tamil | 17.4 | 20.9 | 19.8 | 20.5 |
| Kannada | 11.1 | 13.5 | 13.0 | 11.7 |
| Malayalam | 15.0 | 16.8 | 16.5 | 16.7 |
| Gujarati | 10.2 | 10.5 | 10.1 | 10.6 |
| Odia | 12.5 | 12.9 | 12.3 | 11.1 |
| Marathi | 13.9 | 13.4 | 13.4 | 12.0 |
| Punjabi | 10.9 | 9.4 | 9.1 | 9.7 |
| Assamese | 17.0 | 15.8 | 15.2 | 14.7 |
| Bengali | 12.3 | 10.9 | 10.0 | 10.5 |
| Hindi | 14.2 | 13.2 | 13.1 | 16.6 |
| English | 7.5 | 6.4 | 6.9 | 4.0 |
Read plainly: at well under half the size of the two open 1.2B models, Hima-ASR is roughly level on average and clearly ahead in the four South Indian languages. It trails on Hindi, Bengali and Indian-accented English. On a few clean Hindi sets the error is low (3.4 on Kathbath, 8.8 on Lahaja); on hard conversational sets like SPRING-INX every model is in the 30s to 50s.
The decision head is reported correct on held-out clips such as "wants a human agent" (34 of 34) and language (35 of 36). The card notes the caller demos are acted replies in synthetic voices, so treat those as a sanity check, not a call-centre benchmark.
The data recipe
The model card does not list training data. The recipe comes from a post by the author, shared with us as context. It has two parts: a large pool the author scraped and relabelled, and 34 open datasets.
- Typed by people: 22,121 h (21 sets)
- Read aloud from a script: 13,153 h (4 sets)
- Text from a machine or alignment: 13,787 h (9 sets)
The hours below are as the author reported them. We link a dataset only where we are confident of the official page; search the rest by name on Hugging Face or the publisher's site, and check each license yourself.
Typed by people (22,121 h)
A person listened and wrote down what was said. This is the most trustworthy label and the most expensive.
| Dataset | Hours |
|---|---|
| IndicVoices | 6,835 |
| numo Indic speech | 3,401 |
| SPRING-INX (rounds 1 and 2) | 1,719 |
| Common Voice 22 | 1,559 |
| Hindi-1482 | 1,454 |
| Kathbath | 1,270 |
| Nirantar | 1,242 |
| Gram Vaani phone Hindi | 1,064 |
| Vaani (transcribed) | 941 |
| Equal AI Hindi calls | 652 |
| Rasa expressive | 532 |
| MILE Kannada and Tamil | 377 |
| SUBAK.KO Bengali | 179 |
| Common Voice Indian accent | 164 |
| Rural Women ASR | 153 |
| Indian speech corpus | 120 |
| Malayalam ASR corpus | 116 |
| snorbyte sample | 99 |
| FLEURS | 94 |
| MUCS Hindi | 85 |
| SVQ | 65 |
Read aloud from a script (13,153 h)
A speaker read known text. The label is exact, but the speech is cleaner and flatter than real conversation.
| Dataset | Hours |
|---|---|
| RESPIN | 8,675 |
| HiFiTTS-2 (English audiobooks) | 4,013 |
| SYSPIN | 364 |
| Karya stories | 101 |
Text from a machine or alignment (13,787 h)
The text came from captions, another model, or a script aligned to long audio after the fact. Plentiful, but noisier.
| Dataset | Hours |
|---|---|
| NPTEL lectures | 6,712 |
| Shrutilipi (radio news) | 3,739 |
| OpenBible readings | 940 |
| WordProject Bible | 732 |
| YODAS2 captions | 415 |
| Mann-ki-Baat | 384 |
| UGCE lectures | 346 |
| Spoken-Tutorial | 260 |
| WorldSpeech | 259 |
Why label provenance matters
Grouping by how the text was made, not by language, is the useful idea in this list. A speech model learns whatever its labels say, mistakes included.
Typed
Closest to the truth and to real speech. Use it to anchor the model and to finish training.
Read
Exact text, many speakers, but studio-like. Good for coverage of words and scripts, weak on how people actually talk.
Machine
Huge and cheap, but it carries the errors of whatever produced it. Use it for volume, filter it hard, weight it low.
The 132,130 scraped hours sit in the third bucket too: relabelled means a model wrote the text. That is the standard modern trick (pseudo-labelling), and it is why one engineer can now assemble close to 200,000 hours of training audio. It only works if you filter the machine labels well.
How to train your own, at a high level
This is our sketch of a recipe in the same spirit, not the author's exact pipeline, which is not published. It assumes you have engineers who have fine-tuned a model before and access to a GPU cluster.
- 1Pick a base, do not start from zero. Hima pairs a pretrained speech encoder (Qwen3-ASR-1.7B's) with a small pretrained text decoder (Gemma 3 270M) through a projector. Whisper-style or other open ASR checkpoints work too. Check the base licenses first: a Gemma-derived decoder carries the Gemma Terms of Use into your model.
- 2Check every dataset's license. The 34 sets carry different terms: some allow commercial use, some are research or non-commercial only, some require attribution. Keep a sheet with each set, its license and what it lets you ship. Scraped audio needs its own legal review.
- 3Standardise the audio. Resample everything to 16 kHz mono. Cut long recordings (lectures, radio, Bible readings) into segments under 30 seconds at pauses, keeping the text aligned to each cut.
- 4Normalise the text. One Unicode form (NFC), consistent punctuation and number handling per language, and a decision on code-mixed speech: English words in Latin letters or in the native script. Hima supports both, through
style="mixed". - 5Pseudo-label the unlabelled audio. Run a strong existing model (a "teacher") over raw audio to write transcripts. Drop clips where the teacher is unsure, where two teachers disagree, where the language detected is wrong, or where the text length does not fit the audio length.
- 6De-duplicate. The same audio often appears in several datasets, and Common Voice appears twice in this list. Hash the audio and the text, and remove near-copies.
- 7Keep test sets clean. Several of these sets (FLEURS, Kathbath, SPRING-INX, MILE, Gram Vaani) are also benchmarks. Train only on their train splits, hold out speakers, and never let a test clip or test speaker leak in. Otherwise your WER is flattering and meaningless.
- 8Mix by quality, not just by size. Sample typed data more often than its share of hours, machine-labelled data less. Balance languages, or the biggest sets (IndicVoices, RESPIN, NPTEL) will dominate and the small languages will lag.
- 9Train in stages. First connect encoder and decoder (train the projector, often with the rest frozen). Then train on the full mix. Then finish with a short pass on human-typed data only, which pulls the model back toward clean labels.
- 10Measure per language and per set. Report WER for each language on each test set, plus character error rate (CER) for scripts where word boundaries vary. An average hides weak languages, which is exactly what the Hima card's per-set table avoids.
- 11Optional: add a decision head. If you need intent, urgency or "wants a human", a small head over the encoder and decoder layers is cheaper than a second model. Label real calls for it; synthetic voices are fine for a demo, not for production.
On compute: neither the card nor the post gives a figure, so we will not guess one. What the recipe shows is that the expensive part is no longer the model. It is collecting, filtering and labelling close to 200,000 hours of audio honestly.
Our take
The model is good, but the list is the gift. Most Indic speech work quotes a total hour count; this one says how much of it a person actually typed, and that number (22,121 of about 181,000 hours) is the honest one to plan around. For founders building voice products for India, that means a competitive model is a data-engineering project you can staff, not a research bet. Your edge is the domain audio and labels nobody else has: your callers, your accents, your vocabulary. For the speaking side of the same stack, see our explainer on Maya 2 Native, an open Indic text-to-speech model from Maya Research.
Sources: the Hima-ASR model card on Hugging Face (architecture, runtimes, benchmarks, license), and the dataset breakdown from a post by its author, Prahaladha Reddy. Benchmarks and decision accuracy are self-reported by the author and not independently verified. Dataset hours are as the author reported them. Hima-ASR's own parts and its Qwen3-ASR-derived encoder are Apache 2.0; its decoder is a Gemma derivative, so use and redistribution must follow the Gemma Terms of Use. The training steps are our own sketch. This guide is independent and not affiliated with the author, Maya Research, Google or Alibaba's Qwen team. Spot something out of date? Tell us.