maya-research/maya-transcribe

Maya Transcribe: Hindi & Hinglish Speech-to-Text

Maya Transcribe is an open-weight Hindi and Hinglish speech-to-text model for voice agents and transcription applications. It recognizes Hindi-English code-switching, runs locally with Hugging Face Transformers, and uses an architecture with streaming inference support through a compatible serving backend. The weights are released under Apache-2.0.

5.85% Hindi OI-WER on Voice of India: #3 among 11 models in the open-source filter and #10 of 32 overall, checked on November 3, 2026. The leaderboard lists this release as Maya Transcribe One. View the Hindi STT leaderboard.

In our separate four-model release evaluation, Maya Transcribe ranks #1 by reported macro-average WER: 12.47% across eight Hindi and Hindi-English test conditions. See evaluation methodology and checkpoint-selection details.

It supports two transcription formats: Devanagari, including transliterated English words, or mixed script, with English words in Latin letters. Select the format with style: native or style: mixed.

Quickstart · Hindi voice agents & streaming · Output formats · Benchmarks · FAQ · Credits

Model details

Languages	Hindi and Hindi-English code-switching
Parameters	782.4 million, including the audio encoder
Weights	BF16 · 1.56 GB · Safetensors
Input	16 kHz mono audio, up to 30 seconds per clip
Output	Text in Devanagari or mixed Devanagari–Latin script
Inference	Native Transformers; checkpoint saved with 5.20.2
Streaming	Architecture support via a compatible serving backend; see integration notes below
License	Apache-2.0

The checkpoint stores 782,426,112 parameter elements across 611 tensors, including the audio encoder.

Quickstart

Use Python 3.12 and a fresh environment. Download the model and install the pinned dependencies:

python -m venv .venv
source .venv/bin/activate
pip install -U huggingface_hub
hf download maya-research/maya-transcribe --local-dir maya-transcribe
cd maya-transcribe
pip install -r requirements.txt

Transcribe a local audio file:

python transcribe.py hindi_clip.wav --style native
python transcribe.py hinglish_clip.wav --style mixed --output transcript.json

The CLI loads the local checkpoint, downmixes stereo, resamples to 16 kHz, and uses greedy decoding. It selects CUDA, Apple MPS or CPU when available. Use --device cpu to select CPU explicitly.

Clips longer than 30 seconds must be split before transcription. The CLI rejects overlong input and existing output files. Its JSON output includes a token-limit flag so truncated generation can be detected.

Python

from transformers import AutoProcessor, AutoModelForMultimodalLM

model_id = "maya-research/maya-transcribe"
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=False)
model = AutoModelForMultimodalLM.from_pretrained(model_id, device_map="auto", trust_remote_code=False)

inputs = processor.apply_transcription_request(
    audio="hindi_clip.wav",      # path, URL or a 16 kHz numpy array
    prompt="style: native",      # or "style: mixed" to keep English words in Latin letters
    language="Hindi",
).to(model.device, model.dtype)

output_ids = model.generate(**inputs, max_new_tokens=512, do_sample=False)
generated_ids = output_ids[:, inputs["input_ids"].shape[1]:]
print(processor.decode(generated_ids, return_format="transcription_only")[0])

For a local checkpoint, set model_id to its directory. Always include a style prompt; the model was trained with these prompts to control script choice.

Batch inference

inputs = processor.apply_transcription_request(
    audio=["clip1.wav", "clip2.wav"], prompt="style: mixed", language="Hindi",
).to(model.device, model.dtype)
output_ids = model.generate(**inputs, max_new_tokens=512, do_sample=False)
texts = processor.decode(
    output_ids[:, inputs["input_ids"].shape[1]:], return_format="transcription_only",
)

Adjust batch size to fit your device. See VALIDATION.md for the tested environment and inference checks.

Output formats

Prompt	Format	Example
style: native	Hindi and English words in Devanagari	पेमेंट पेंडिंग है
style: mixed	Hindi in Devanagari; English words in Latin script	payment pending है

The examples above illustrate formatting. Use the same convention as your reference transcripts when evaluating recognition accuracy.

Audio example

The following 7.76-second clip is synthetic speech generated by Maya 2 Native. Both excerpts are actual outputs from this checkpoint:

Prompt	Transcript excerpt
style: native	पैंतालीस एमजी रोड अंधेरी ईस्ट मुंबई
style: mixed	पैंतालीस mg road अंधेरी east मुंबई

Full transcripts and test environment.

Evaluation

Voice of India Hindi STT leaderboard

The independent Voice of India speech-to-text leaderboard reports the following Hindi results for Maya Transcribe One, checked on November 3, 2026, with language set to Hindi and all speaker/audio filters left at their defaults:

Metric	Result
Hindi OI-WER	5.85%
Rank in the open-source filter	3 / 11
Rank across all licences	10 / 32
Hindi evaluation samples	21,020

The two models above Maya in the open-source filter are Bodhan's Indic Transcribe Core and Flex. Maya scores ahead of Nemotron 3.5 Transcribe, SraVaani, Gemma E4B, OmniASR LLM 7B v2 and Qwen3-ASR on this Hindi board.

Voice of India's methodology uses spontaneous telephonic speech and orthography-invariant word error rate (OI-WER), allowing accepted spelling variants. This is a different evaluation and metric from the release WER results below. The leaderboard measures transcription accuracy, not streaming behaviour or latency; its rankings are not a streaming-specific ranking. Live positions may change.

Release evaluation

The release evaluation covers 12,923 clips and 173,680 reference words across eight Hindi test conditions. Results are normalized word error rate (WER, %); lower is better.

All four systems were scored on the same clips. The normalizer removes punctuation and case, and treats split and joined abbreviations such as ई पी एफ ओ and ईपीएफओ as equivalent. Maya Transcribe uses greedy decoding, with style: mixed for the code-switching set and style: native elsewhere.

Checkpoint selection: half of each test set was used to choose the checkpoint. The full-set table includes those samples. Scores on the half excluded from checkpoint selection are listed separately below.

Test set	Clips	Maya Transcribe	indic-transcribe-core	indic-transcribe-flex	Sarvam Saaras v3 (API)
Kathbath	1,929	9.36	10.38	10.74	9.61
Kathbath noisy	1,929	10.01	10.95	11.25	10.48
FLEURS	418	10.72	10.10	10.35	10.01
Common Voice	1,727	11.54	11.01	11.57	12.02
MUCS	3,897	9.67	9.53	10.32	9.72
Lahaja (accents)	503	11.50	10.26	10.49	13.94
Code-mix (Hindi-English)	1,488	16.97	43.03	17.05	20.69
Gram Vaani (phone calls)	1,032	19.94	18.48	18.82	22.95
Reported macro-average	12,923	12.47	15.47	12.57	13.68

Maya Transcribe has the lowest reported macro-average in this comparison and the lowest WER on Kathbath, Kathbath noisy and Hindi-English code-switching. It trails other systems on the remaining five conditions. The 0.10-point macro-average difference from indic-transcribe-flex should not be treated as a statistically significant lead.

Reported macro-WER on the half not used for checkpoint selection	WER
Maya Transcribe	12.28
indic-transcribe-core	15.24
indic-transcribe-flex	12.46
Sarvam Saaras v3 API	13.79

Evaluation notes:

• Macro-average gives each test condition equal weight. It is not pooled, word-weighted WER. Values are reported from the release evaluation; recomputing from rounded table entries may differ slightly.
• Code-switching references use Latin-script English. Transcribing those words in Devanagari increases WER even when the spoken word is recognized.
• Nine Gram Vaani clips returned no Sarvam transcript and were counted as fully incorrect for that system.
• Dataset versions, exact split files, comparator revisions, per-clip outputs and the scoring implementation are not included. The evaluation has not been independently reproduced from those artifacts.

Training

All model parameters were updated on Hindi and Hindi-English speech using next-token transcript loss. The released weights are an average of selected checkpoints from the final training run.

Training hours, dataset identities, detailed data provenance, deduplication, hyperparameters and checkpoint-averaging settings have not been published.

Hindi voice agents and streaming

For developers building Hindi voice AI, Hinglish assistants, live transcription or call-centre voice agents, Maya combines downloadable weights, script control and a streaming-capable model architecture.

• Hindi-English code-switching: use style: mixed to retain English terms in Latin script for downstream language models, search and transcript review.
• Self-hosted speech recognition: run the downloaded checkpoint in your own environment under Apache-2.0.
• Streaming integration: incremental audio ingestion and partial transcripts belong to the serving backend. The included transcribe.py handles complete clips; it does not start a microphone or WebSocket service.

The upstream architecture has a documented streaming inference implementation. Current vLLM documentation also describes a real-time ASR path through /v1/realtime. These are backend references, not a tested deployment recipe for this native Transformers checkpoint. Check checkpoint-format compatibility, preserve the style: native or style: mixed system prompt, and validate partial transcripts, final transcripts and end-to-end latency before integration.

For comparison, Bodhan's Indic-Transcribe release article describes its current models as recorded-audio transcription and lists real-time streaming as future work, as checked on November 3, 2026. Maya's streaming-capable architecture is relevant when selecting a Hindi ASR model for a live voice-agent stack. There is no published streaming-mode accuracy or latency benchmark for this Maya checkpoint yet.

Serving

The release notes report vLLM 0.29 serving on one A100 40 GB: approximately 0.16 seconds per single-request clip and 76 clips/second at concurrency 32. Clip durations, latency clock, warm-up and percentile measurements are not documented, so these figures are not deployment capacity targets.

The included inference script uses Transformers. Native and mixed-script transcription, two-item batch inference, and seven CLI checks passed on Apple Silicon with MPS. This does not validate the vLLM/A100 serving figures. Details are in VALIDATION.md.

Limitations

• Evaluation covers Hindi and Hindi-English speech. English-only speech and other languages were not evaluated.
• Phone-call audio remains difficult: Gram Vaani WER is 19.94%.
• Split recordings longer than 30 seconds. This repository does not include diarization, word timestamps, VAD-based long-form transcription or a validated real-time streaming service.
• Names, numbers, abbreviations and code-switched words may have multiple valid written forms. Check important transcripts against the recording.

Process audio you have permission to use and protect recordings and transcripts appropriately.

Developer FAQ

Can I use Maya Transcribe for Hindi speech recognition and Hinglish transcription?

Yes. The release focuses on Hindi and Hindi-English code-switching. Set language="Hindi" and use style: native for Devanagari output or style: mixed for Hindi with English words in Latin script. English-only speech and other languages have not been evaluated for this release.

Does Maya Transcribe support streaming ASR for voice agents?

Its model architecture supports streaming inference through a compatible serving backend. This repository provides complete-clip and batch inference. See streaming integration notes for backend references and the validation needed for this checkpoint.

Is Maya Transcribe the best Hindi ASR model?

It has the lowest reported macro-WER among the four systems in our release evaluation. On the independent Voice of India Hindi leaderboard, it ranks #3 in the open-source filter and #10 overall as of November 3, 2026. These results do not establish a universal or streaming-specific #1 rank. Choose using accuracy, script format and measured latency on your own audio.

Can I use it commercially and run it without a hosted API?

Yes. The model weights are Apache-2.0 and can be downloaded for local inference. After downloading the dependencies and checkpoint, local transcription does not require a hosted Maya API. Preserve the licence and attribution when redistributing.

License

Apache-2.0. Preserve the included NOTICE when redistributing.

Credits

Developed by Maya Research through full-parameter fine-tuning of Qwen/Qwen3-ASR-0.6B-hf. We thank the Qwen team for the pretrained model and Hugging Face and the Transformers contributors for the native inference implementation. The upstream “0.6B” family name refers to the decoder scale; the full checkpoint includes the audio encoder.

@misc{maya_transcribe_2026,
  title = {Maya Transcribe: Hindi and Hindi-English Speech Recognition},
  author = {{Maya Research}},
  year = {2026},
  url = {https://huggingface.co/maya-research/maya-transcribe}
}