Every amount right
Reads pound amounts, card digits and reference codes so customers hear them right.
100% heard rightBenchmarks
Built for UK customer service.
Every amount heard right.
Word error rate by voiceLower is better · word error rate %
3.87Base model
3.87%
3.87%
1.62Balsa Voice v1
1.62%
1.62%
4.12Base model
4.12%
4.12%
1.5Balsa Voice v1
1.5%
1.5%
3.61Base model
3.61%
3.61%
2Balsa Voice v1
2%
2%
Every check, in detail21 checks
| Check | Base model | Balsa Voice v1 | v1 against base |
|---|---|---|---|
| Times real time (xRT), MacBook Air M5How many times faster than real time it speaks | 21.6× | 21.6× | = |
| RTF p50 and p90, every clip, MacBook Air M5Typical and slow clips, model load included | 0.046 and 0.052 | 0.046 and 0.052 | = |
| One hour of audio, MacBook Air M5Time to make an hour of speech | About 3 minutes | About 3 minutes | = |
| Slowest clip, times real time (xRT)Speed of the slowest clip | 16.7× | 16.7× | = |
| Peak memory, MacBook Air M5Most memory used while speaking | 253 MB | 253 MB | = |
| Deterministic output, same seedSame text and seed give the same file | 100% | 100% | = |
| ASR WER, all voicesWords a speech-to-text model got wrong, all voices | 3.87% | 1.62% | −2.25 pts |
| ASR WER, EleanorWords a machine misheard in the Eleanor voice | 1.03% | 1.50% | +0.47 pts |
| ASR WER, HarrietWords a machine misheard in the Harriet voice | 4.12% | 1.50% | −2.62 pts |
| ASR WER, OliverWords a machine misheard in the Oliver voice | 3.61% | 2.00% | −1.61 pts |
| ASR WER, RhysWords misheard in the Rhys voice, not shipped | 6.70% | 1.50% | −5.20 pts |
| Wrong words (substitutions)Words the machine heard as a different word | 16 | 6 | −10 |
| Missed words (deletions)Words the machine did not hear at all | 12 | 6 | −6 |
| Added words (insertions)Extra words the machine heard | 2 | 1 | −1 |
| Text normalisation, pound amountsPound amounts heard back exactly | 100% | 100% | = |
| Text normalisation, card digitsCard digits heard back in the right order | 100% | 100% | = |
| Text normalisation, reference codeReference code heard back letter for letter | 100% | 100% | = |
| Names pronounced rightNames such as Leeds and Cardiff heard right | 100% | 100% | = |
| Text normalisation, datesDates such as 1 December spoken the UK way | 4 of 8 | 8 of 8 | +4 |
| Time format errors (text normalisation)Times like 8 a.m. heard in the wrong format | 24 | 0 | −24 |
| Domain-term pronunciation flagsBanking words a machine did not hear back | 13 of 160 | 10 of 160 | −3 |
Measured on multiple UK customer service sentences in multiple voices. WER comes from a speech-to-text round trip, scored by an open speech-to-text model after text normalisation. The base model is an open model under Apache 2.0, run before the Balsa text step, which spells out UK dates and times and says cheque the UK way. Speed counts model load on every call, over three runs on a MacBook Air M5. How we test
Capabilities
One model,
four jobs done well.
Faster than real time
Makes an hour of audio in about 3 minutes on a MacBook Air M5.
21.6× real timeSays it the UK way
The Balsa text step spells out UK dates and times, and says cheque, not check.
0 time format errorsRuns on device
A 0.04B model that needs 253 MB of memory at peak. No cloud.
103 MB fileModel card
0.04B parameters.
21.6× real time, on device.
Balsa Voice
A 0.04B text-to-speech model for the UK. It reads customer service messages in British English on device, 21.6 times faster than real time on a MacBook Air M5.
- Built for
- The UK
- Built on
- An open model (Apache 2.0)
- Base model languages
- English, as published by its maker
- Weights
- 103 MB file
- Domain
- General use, in British English. Tested on UK customer service messages.
- Version
- v1, November 2026
Spec
- Model
- balsa-voice-0.04b-v1
- Parameters
- 0.03B, 0.04B with the audio tokenizer (31M and 41M)
- File
- 103 MB
- Memory
- 253 MB peak
- Language
- British English
- Voices
- Eleanor, Harriet, Oliver
- Output
- Speech, 24 kHz mono
- Runs on
- Apple silicon, on device
Times faster than real time, MacBook Air M5
21.6×
Median of three runs, with model load counted on every call. An hour of audio takes about 3 minutes.
Repeatability
48/48
Every clip came out byte-identical across three runs with the same seed.
Built for the UK
- Pound amounts heard right100%
- Card digits and reference codes heard right100%
- Faster than real time, M5 Air21.6×
- Same seed, same audio100%
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 Balsa models
Five models,
each doing one job.
Balsa Omni
Balsa Omni v1 is a multimodal model for the UK, with 4.5B effective parameters. It reads text, images and audio, replies in British English and runs on device, with no cloud.
4.5BView modelBalsa TextBalsa Text v1 is a small assistant model for the UK. It answers in British English and runs on device, on a laptop, with no cloud.
2.3BView modelBalsa VisionBalsa Vision v1 turns scanned English pages into text for the UK, on device with no cloud. It reads 98.5% of characters right on pages 900px and wider.
0.3BView modelBalsa VoiceBalsa Voice v1 reads UK customer service messages aloud in British English, on device with no cloud, 21.6 times faster than real time on a MacBook Air M5.
0.04BYou are hereBalsa DecisionBalsa Decision v1 flags deceived payments and routes UK customer messages, on device with no cloud. It recommends an option, and a person decides.
0.3BView modelQuestions
Asked often.
Balsa Voice v1 is a 0.04B (41M parameter) text-to-speech model for the UK. It reads customer service messages aloud in British English, with pound amounts, dates, times, card digits and reference codes. It is built on an open model, under the Apache 2.0 licence.
No. Balsa Voice v1 runs on device. On a MacBook Air M5 with 16 GB, it uses 253 MB of memory at peak.
Balsa Voice v1 speaks 21.6 times faster than real time on a MacBook Air M5, with model load counted. An hour of audio takes about 3 minutes to make.
Balsa Voice v1 speaks British English in three voices: Eleanor, Harriet and Oliver. Balsa models are built for the UK.
Balsa Voice v1 was tested on the hardest parts of a UK customer service message: pound amounts, dates, times, card digits and reference codes. Every clip was checked by ear and by a speech-to-text model. A score of 100% means every clip passed that check.
Write to research@f-r.co.
