Skip to the results
ampixa
hear// Rani · Nepali speech recognition

A phone call is
a different test.

Picture someone in Nepal calling a support line. They need their words written down. A quiet microphone is a different test. NepaliConformer is now Rani.

Raniरानी
33.5word errors per 100 words · NepTel 0.2Public call test
30.6word errors per 100 words · private calls, machine referencesInternal call test
3.6letter errors per 100 letters · PVSpeech, with a language modelRead speech test
// try it

Speak Nepali. Read it back.

Record or upload up to 60 seconds. It runs on a free CPU, so it may take a minute to wake up. Open full-screen if your microphone is blocked.

On the line

Start with the call.

Rani turns spoken Nepali into text. That can help with helplines, captions, voice typing and meeting notes. But a recogniser useful in a quiet room can fail on a call. We tested both. On NepTel, Rani leads the ranked systems shown here without a language model. Each lead has a 95% confidence interval excluding zero. That supports a lead on this sample, not on every call.

The private test uses real Nepali customer-service phone calls: 127 calls Rani never trained on. Rani leads these comparators too. The references are machine-made by Scribe and cannot be checked publicly.

NEPTEL 0.2 · mistakes per 100 words
Lower is better. Rani uses no language model.
Rani33.5
NepaliConformer offlineour older model39.7
AI4Bharat IndicConformer47.6

Human-corrected references. Every reference was corrected against the audio; the drafting model is unranked. Results and reference provenance.

PRIVATE CALLS · mistakes per 100 words
Internal evaluation. Rani uses no language model.
Rani30.6
AI4Bharat IndicConformer38.9
NepaliConformer offlineour older model49.2

Machine-made Scribe references; not public ground truth. Unseen calls. Each lead has a 95% interval excluding zero. Internal evidence and limits.

// the answer key changed
Chirp 2
old draft source · NepTel 0.1
A person listened.
every reference corrected · NepTel 0.2
against the audio

The answer key is part of the test.

NepTel’s old labels were drafted by Google’s ASR, Chirp 2, then reviewed. In NepTel 0.2, a human listened to and corrected every reference against the audio. A draft can carry its model’s mistakes into the answer key, so the drafting model is not ranked.

Read speech.

Clean speech, and a language model.

Mistakes per 100 letters. Rani with a language model: 3.6 on PVSpeech and 10.4 on FLEURS Nepali. AI4Bharat’s IndicConformer: 9.8 and 7.4.

Mistakes per 100 letters · lower is better
ModelPVSpeechFLEURS
Rani (language model)3.610.4
AI4Bharat IndicConformer9.87.4

PVSpeech: selection set, nonpublic.
Language model not in the released weights. Clean-speech receipts.

Speed.

How fast does it write?

On one CPU thread, Rani transcribes 26.8 times faster than real time. A 10-second clip returns text 0.35 seconds after it ends. It decodes offline, so text arrives when the clip ends, not while you speak.

One CPU thread · Apple M2 Ultra
Model× real timeFirst text, 10 s clip
Rani26.8×0.35 s
NepaliConformer offline27.1×0.33 s
AI4Bharat IndicConformer21.1×0.41 s
Rani-IndicConformer20.8×0.38 s

Speed receipts.

Time to first text: a 10-second clip returns text 0.35 seconds after it ends on one CPU thread
// same family, different jobs

What changed with the name?

The family has three classes. NepaliConformer offline is the older model. Rani continues that line, trained further for customer calls and read speech. Rani-IndicConformer is a full IndicConformer fine-tune aimed at clean speech. On our internal PVSpeech measurement, it cuts letter errors from IndicConformer’s 9.8 to 3.9 per 100 letters.

NepaliConformer offline
The older model.
The public downloads are this earlier work.
Rani
The continued model line.
Customer calls and read speech.
Rani-IndicConformer
A full IndicConformer fine-tune.
Clean speech.

The older model’s encoder, which turns sound into useful representations, was initialised from AI4Bharat’s MIT-licensed IndicConformer Nepali encoder. The decoder and Nepali vocabulary are ours, trained on our data. Rani continues this lineage. It uses a TDT decoder, predicting text tokens and their durations, and a Nepali-only vocabulary. The scores measure these choices together, not each choice’s individual benefit.

// weights, with approval

The next test should be yours.

All flavours in one place.
Approval required to download.

Rani, Rani-IndicConformer and the older NepaliConformer models are together on Hugging Face, and downloading requires approval. Distillers must visibly credit Ampixa Labs in their model card and documentation. The licence also covers training or evaluating another model with our outputs and building or distributing derivatives; the requirement passes onward. Read it before building. Next, use the public NepTel audio and scorer to check your system on the same test. NepTel is for evaluation; don’t train on it.

Request the weightsRead the repositoryRead the model licence
// receipts · every visible number

The number belongs to the test.

Public results link to their source. Private call results are internal measurements and are marked as such.