AksharSTT Lite v4

Offline speech-to-text for Hindi, English and Hinglish, in a 20.4 MB file for phones and devices with little memory.

Website · Live demo · GitHub · AksharSTT v4 · AksharSTT v1

AksharSTT Lite v4 is AksharSTT v4 retrained at lower precision: most of its decoder is 2-bit, and its encoder stays 4-bit. It is 9.1 MB smaller than AksharSTT v4 and 0.93 points of mean word error rate behind it. Like AksharSTT v4, it runs in Cactus's needle engine on the device itself.

Highlights

• 20.4 MB, against 29.5 MB for AksharSTT v4 and for Shruti at 4 bits.
• Mean word error rate 21.71 over six test sets: 0.93 behind AksharSTT v4 and 0.07 below Shruti at 4 bits (21.78), on the same clips, scored by the same harness.
• Indian-accented English (Svarah): 22.43, level with Shruti (22.45) and 2.25 better than AksharSTT v1.
• Lower than AksharSTT v1 on five of the six sets and on the mean (21.71 against 22.28).
• On 100 clips of silence and faint noise it writes nothing.

All scores are on the full test sets with the settings given under Accuracy.

Model details

Quick start

pip install cactus-needle==3.2.8 huggingface_hub

import os
os.environ.setdefault("NEEDLE_TELEMETRY", "0")  # needle reports usage unless these are set
os.environ.setdefault("DO_NOT_TRACK", "1")

from huggingface_hub import hf_hub_download
from needle import Whistle

model = Whistle(hf_hub_download("byteknotlabs/AksharSTT-Lite-V4", "akshar-lite-v4.cact"))
print(model.transcribe("speech.wav", language="hi")["text"])

Use language="hi" for Hindi and Hinglish, and language="en" when you know the speech is English (see Language flag). transcribe takes a WAV file or 16 kHz mono float samples in [-1, 1], of any length.

With word_timestamps=True the result also lists every word with its start and end in seconds and the model's probability for it:

result = model.transcribe("speech.wav", language="hi", word_timestamps=True)
for w in result["words"]:
    print(w["word"], w["start"], w["end"], w["probability"])

Output

The model writes one style: Hindi in Devanagari, English words in lowercase Latin letters, numbers as words, and no punctuation. On two test clips it writes:

The training transcripts write English words, brands and place names in Latin when they are on a curated list of 2,900 words (data/english-words), so the model does the same; other loanwords can come out in either script. Punctuation, digits and other scripts are left to the application.

Accuracy

Word error rate in percent on the full test sets, lower is better.

All four models were scored on the same clips and the same machine (AMD Ryzen 9 7900X) by akshar's harness, python -m akshar.eval: normaliser nfc-lower-punct2space-v1, needle engine build 3.3.3 (cactus-needle 3.2.8) with its default 5 beams, language flag hi for Monsoon, FLEURS, Lahaja and MUCS and en for Svarah and LibriSpeech. Monsoon is scored with OIWER, which accepts any of the spellings its references list for a word, Devanagari or Latin; the other sets with plain WER. On 100 clips of silence and faint noise none of the four models writes a word. The protocol and the pinned test-set revisions are in docs/eval.md.

Against Shruti at 4 bits, AksharSTT Lite v4 is ahead on FLEURS (by 1.07), Lahaja (by 0.51) and MUCS (by 0.41), level on Svarah, and behind on Monsoon (by 1.01) and LibriSpeech (by 0.55).

Language flag

The flag matters for English speech. With hi, AksharSTT v4 scores 26.97 on Svarah; with en, 21.41 (same clips, same harness and settings otherwise). The lite model was not scored with hi on Svarah. Apps should pass en when they know the speech is English, and hi for Hindi and Hinglish.

Training

How it was made

• Start: AksharSTT v4's final full-precision weights, the ones published as akshar-v4.safetensors in byteknotlabs/AksharSTT-V4.
• Precision: quantisation-aware training at the target widths: Cactus's 2-bit mix (embedding table and routing matrices at 4 bits, the other matrices at 2 bits) with the encoder raised to 4 bits.
• Teacher: AksharSTT v4 as released (4-bit). The loss is cross-entropy on the targets plus distillation from the teacher at weight 0.5, temperature 1.0.
• Run: 3,600 updates at a learning rate of 4e-4, about 18 minutes on one RTX 5080. Peak GPU memory 4.7 GB. An update is about 480 seconds of padded audio.
• Validation cross-entropy on 2,691 held-out clips: 0.3309 just after the switch to 2 bits, 0.2339 at the end (AksharSTT v4: 0.2111).
• Data and targets: the same as AksharSTT v4's, from the same stream and the same 230 hours kept on disk. This run drew 406 hours from the stream.

AksharSTT v4 itself was trained from Cactus Whistle, with Shruti as its teacher and no weights taken from Shruti. Its card describes the lineage, the changes in v4 and the data in full. The record of this run is training/lite-v4.json.

Data

AksharSTT Lite v4 carries what AksharSTT v4 learned, so its data is v4's:

Hours count each clip once: about 3,930 in AksharSTT v4's run and about 640 in this one. Every training clip that matches a test clip by its audio, or by its text when the text has three or more words, was removed.

Intended use

• Transcribing Hindi, English and Hinglish speech without a network where memory or download size is tight: phones, small devices and apps that bundle the model.
• Apps that must keep audio on the device.

Where size allows, AksharSTT v4 is more accurate on every test set.

Out of scope: languages other than Hindi and English, punctuated or formatted text, translation, and telling speakers apart.

Limitations

• Against AksharSTT v4: 0.93 points of mean WER behind, most on phone conversations (Monsoon, by 2.08) and Indian English (Svarah, by 1.02).
• Against Shruti at 4 bits: behind on Monsoon (by 1.01) and LibriSpeech (by 0.55).
• Hinglish tutorials (MUCS): 34 to 35 WER for every model in the table. MUCS references are the narrators' scripts rather than verbatim transcripts, and they write the same English word in either script, so part of that error is in the references.
• Accented Hindi (Lahaja): 25.78, 1.25 worse than the internal lite v3, the same pattern as AksharSTT v4 (see its card).
• Script on FLEURS and Lahaja: their references write English loanwords and place names in Devanagari, so a correct word written in Latin counts as an error on those two sets.
• Numbers and punctuation: numbers come out as words, with no punctuation or casing.
• Languages: trained and tested on Hindi and English only. Whistle's other language tags are still in the vocabulary, untested.
• Engine build: the scores were measured with engine build 3.3.3 on x86. Other builds and other processors can move them a little.
• Devices: speed and memory on phones have not been measured yet.
• Clip length: trained on clips of up to 30 seconds. The engine accepts longer audio, but long recordings were not scored on their own.

Responsible use

Transcripts from this model are machine-generated and can be wrong. Label them as machine-generated when you show them to people or pass them on. The Mozilla Data Collective terms for Common Voice, one of the training sources, ask that outputs of models built with that data be disclosed to downstream users as machine-generated. Have a person check a transcript before it is used in a decision about someone.

Files

sha256 of akshar-lite-v4.cact: 9a7bdf9be822d66758a13ad0ec9ef90c6ec47a88a9be1aeb526ba9d3207eff6b

The full-precision weights it was made from are in byteknotlabs/AksharSTT-V4.

Licence

Apache-2.0. AksharSTT v4, Whistle and Shruti are Apache-2.0 too. The training datasets are CC-BY-4.0, except MUCS 2021, which is CC-BY-SA-4.0, and Common Voice, which is CC0. NOTICE credits each of them.

Citation

@software{aksharstt2026,
  title   = {AksharSTT: offline speech-to-text for Hindi, English and Hinglish},
  author  = {{ByteKnot Labs}},
  year    = {2026},
  version = {Lite v4},
  url     = {https://huggingface.co/byteknotlabs/AksharSTT-Lite-V4},
  note    = {Code: https://github.com/byteknotlabsai/akshar}
}