Every rupee amount
Speaks Hindi and Marathi amounts, dates and digits so they are heard back right.
6/6 heard back rightBenchmarks
Built for India payments.
Faster than speech.
Error rate by languageLower is better · error rate %
13.95Base model
13.95%
13.95%
12.79Gati Voice v1
12.79%
12.79%
4.33Base model
4.33%
4.33%
3.94Gati Voice v1
3.94%
3.94%
4.29Base model
4.29%
4.29%
2.86Gati Voice v1
2.86%
2.86%
5Base model
5%
5%
3.89Gati Voice v1
3.89%
3.89%
6.09Base model
6.09%
6.09%
6.09Gati Voice v1
6.09%
6.09%
Every check, in detail25 checks
| Check | Base model | Gati Voice v1 | v1 against base |
|---|---|---|---|
| RTF, MacBook Air M5 (lower is better)Seconds of work per second of speech, M5 Air | 0.83 | 0.76 | −0.07 |
| RTF p50 and p90, every clip, MacBook Air M5Typical and slow clips across three back-to-back runs | Not run | 0.78 and 0.80 | |
| Times real time, MacBook Air M5How many times faster than speech it runs | 1.20× | 1.32× | +0.12× |
| One hour of audio, MacBook Air M5Time to make an hour of speech | About 50 minutes | About 46 minutes | −4 minutes |
| Clips faster than real time, three runs back to backClips with a real-time factor below 1.0 | Not run | 84/84 | |
| Slowest clip, MacBook Air M5Real-time factor of the slowest of 84 clips | Not run | 0.83 | |
| RTF, iMac M1 (lower is better)Seconds of work per second of speech, iMac M1 | 2.09 | 1.74 | −0.35 |
| Time to first audio (TTFA), MacBook Air M5Wait before the first sound plays, M5 Air | 0.34 s | 0.36 s | +0.02 s |
| TTFA p50 and p90, every clip, MacBook Air M5Typical and slow wait before the first sound | Not run | 0.35 s and 0.37 s | |
| Time to first audio (TTFA), iMac M1Wait before the first sound plays, iMac M1 | 0.91 s | 0.75 s | −0.16 s |
| Peak memory, MacBook Air M5Most memory used while speaking, M5 Air | 6.08 GB | 5.87 GB | −0.21 GB |
| Peak memory, iMac M1Most memory used while speaking, iMac M1 | 6.36 GB | 6.70 GB | +0.34 GB |
| Listener acceptance, HindiHindi clips a listener would play to a customer | Not run | 100% | |
| Native speaker acceptance, MarathiMarathi clips a native speaker would play | Not run | 100% | |
| Deterministic output, same seedSame text and seed give the same file | Not run | 100% | |
| ASR WER, HindiHindi words a speech-to-text model got wrong | 12.39% | 14.22% | +1.83 pts |
| ASR WER, MarathiMarathi words a speech-to-text model got wrong | 13.95% | 12.79% | −1.16 pts |
| ASR WER, EnglishEnglish words a speech-to-text model got wrong | 4.29% | 2.86% | −1.43 pts |
| ASR WER, overallWords got wrong across all three languages | 11.23% | 11.76% | +0.53 pts |
| ASR CER, HindiHindi characters a speech-to-text model got wrong | 7.39% | 8.00% | +0.61 pts |
| ASR CER, MarathiMarathi characters a speech-to-text model got wrong | 4.33% | 3.94% | −0.39 pts |
| ASR CER, EnglishEnglish characters a speech-to-text model got wrong | 5.00% | 3.89% | −1.11 pts |
| ASR CER, overallCharacters got wrong across all three languages | 6.09% | 6.09% | = |
| Text normalisation, rupee amounts and a dateRupee amounts and a date survive the round trip | 100% | 100% | = |
| Text normalisation, six digits in orderA six-digit number heard back in order | 100% | 100% | = |
The base model is an open model under Apache 2.0, run here at full precision. Gati Voice v1 has an 8-bit backbone. Both ran the same UPI payment sentences in multiple voices, on the same machines. WER and CER come from a speech-to-text round trip, scored by open speech-to-text models after text normalisation. Speed is the median of three back-to-back runs on a fanless MacBook Air M5, and p50 and p90 cover every clip in those runs. How we test
Capabilities
One model,
four jobs done well.
Faster than speech
Speaks faster than speech on a MacBook Air M5, with first audio in 0.36 s.
0.76 real-time factorHindi and Marathi
Four built-in voices for Hindi, Marathi, Hinglish and Indian English.
10/10 Hindi, 8/8 MarathiRuns on device
A 2.5B model with an 8-bit backbone. It needs 5.87 GB of memory at peak, with no cloud.
3.31 GB fileModel card
2.5B parameters.
Runs on a laptop.
Gati Voice
A 2.5B text-to-speech model for India. It speaks Hindi, Marathi, Hinglish and Indian English on device, faster than speech on a MacBook Air M5.
- Built for
- India
- Built on
- An open model (Apache 2.0)
- Base model languages
- 12 languages, including Hindi, Marathi and English, as published by its maker
- Audio codec
- Mimi, by Kyutai (CC BY 4.0)
- Domain
- General use, in Hindi, Marathi, Hinglish and Indian English. Tested on UPI payment messages.
- Version
- v1, November 2026
Spec
- Model
- gati-voice-2.5b-v1
- Parameters
- 2.5B
- Weights
- 8-bit backbone, MLX
- Memory
- 5.87 GB peak
- Languages
- Hindi, Marathi, Hinglish, Indian English
- Voices
- Aanya, Kabir, Gauri, Omkar
- Output
- Speech, 24 kHz mono
- Runs on
- Apple silicon, on device
Real-time factor, MacBook Air M5
0.76
Lower is faster. Below 1 is faster than speech. Across 84 clips in three back-to-back runs, no clip reached 1.0.
Repeatability
28/28
Every clip came out byte-identical across three runs with the same seed.
Built for India
- Hindi clips accepted100%
- Marathi clips accepted, native speaker100%
- Amounts and dates heard back right100%
- Faster than speech, M5 Air0.76
Film
Watch it run
on device.
How it speaks.
- 01Type a line
- 02Read the numbers
- 03Speak
- 04Stream as it goes
Playground
Try it yourself.
Watch it work, step by step.
Pick a step or let it play. Demos tagged as illustrations show example output, and the scores on this page come from our own tests.
More Gati models
Five models,
each doing one job.
Gati Omni
Gati Omni v1 is a multimodal model for India, with 4.5B effective parameters. It reads text, images and audio, replies in eight languages used in India and runs on device.
4.5BView modelGati TextGati Text v1 is a small assistant model for India. It answers in Hindi, Marathi, Gujarati, Telugu, English and Hinglish, on device, with no cloud.
2.3BView modelGati VisionGati Vision v1 turns scanned Hindi, Marathi and English pages into text for India, on device with no cloud. It reads best on sharp scans 900px and wider.
1BView modelGati VoiceGati Voice v1 reads UPI payment messages aloud in Hindi, Marathi, Hinglish and Indian English, on device, faster than speech on a MacBook Air M5.
2.5BYou are hereGati DecisionGati Decision v1 reads Hindi payment messages for India and recommends the first step, on device with no cloud. A person makes the final decision.
0.6BView modelQuestions
Asked often.
Gati Voice v1 is a 2.5B text-to-speech model for India. It reads UPI payment messages aloud, with rupee amounts, dates and digits. It is built on an open model, under the Apache 2.0 licence.
No. Gati Voice v1 runs on device. On a MacBook Air M5 with 16 GB, it speaks faster than speech and uses 5.87 GB of memory at peak.
Gati Voice v1 runs live on a MacBook Air M5 with 16 GB, at a real-time factor of 0.76. Below 1 is faster than speech. On an iMac M1 the figure is 1.74, so use it there for audio made ahead of time.
Gati Voice v1 speaks Hindi, Marathi, Hinglish and Indian English in four built-in voices: Aanya, Kabir, Gauri and Omkar. Gati models are built for India.
Gati Voice v1 was tested on the hardest parts of a payment message: rupee amounts, dates, digits and code-mixed Hindi and English. Every clip was checked by ear, and the Hindi, Marathi and English clips by a speech-to-text model as well. A score of 100% means every clip passed that check.
Write to research@f-r.co.
