New speech recognition model
Gaeilge

Irish speech to text. The most accurate in GASTON

Our own speech-to-text model for Irish, trained on about 1,890 hours of Irish speech. It turns radio, podcasts, parliament debates and interviews into searchable text, with timestamps for every word.

25.3%
word error rate on parliament speech
12.1%
word error rate on read speech
~1,890 h
of Irish speech in training
~½
the word errors of ElevenLabs Scribe v2

Word error rate counts the words that differ from a careful human transcript. Lower is better.

01 · Benchmark

About half the word errors of ElevenLabs Scribe v2

We ran ElevenLabs Scribe v2, set to Irish, and BuzzASR, the public Irish fine-tune of OpenAI Whisper, on the same four public test sets and scored all of them the same way. Averaged over the four sets, our model gets 24.3% of words wrong, Scribe v2 46.2%.

The gap is widest on parliament speech and on dialect recordings. Stock OpenAI Whisper has no Irish option at all: forced to transcribe Irish, its error rate is above 100%.

Word error rate on four public test sets
Lower is better
Gaston Irish ElevenLabs Scribe v2 BuzzASR
Oireachtas · Parliament committees
25.3%
46.9%
51.9%
Canúint · Dialect archive, older recordings
27.7%
58.8%
66.3%
FLEURS · Read Wikipedia sentences, many speakers
32.1%
43.4%
49.0%
Common Voice · Short read sentences
12.1%
35.6%
n/a*

Tested in November 2026 on 4,680 clips (about 14 hours), the full public test sets. Scribe v2 ran through the ElevenLabs API with the language set to Irish. 95% confidence intervals for our lead over Scribe v2: 21 to 22 points (Oireachtas), 29 to 34 (Canúint), 10 to 12 (FLEURS), 22 to 25 (Common Voice). Commercial models change over time.

* BuzzASR is not scored on Common Voice: its training data contains the sentences of that test set.

02 · Accuracy

Accuracy by kind of audio

Besides the four public test sets, we measure two harder, non-public ones: Irish YouTube videos and spontaneous conversation recorded in Mayo. Those two stay in-house, so only our own model is scored on them.

Common Voice
Short read sentences
12.1%
WER
YouTube
Vlogs, interviews, multi-speaker video
21.3%
WER

Scored against partly condensed subtitles, so this figure is too high.

Oireachtas
Parliament committees
25.3%
WER
Canúint
Dialect archive, older recordings
27.7%
WER
FLEURS
Read Wikipedia sentences, many speakers
32.1%
WER
abairt.com
Spontaneous conversation, Mayo dialect
49.8%
WER

The hardest kind of Irish audio we test.

Word error rate in percent, on normalised text, decoded with beam search 5.

03 · What to expect

Best on clear speech that is mostly Irish

News, radio, podcasts, lectures, parliament, read text

A usable transcript. Roughly 1 word in 5 to 8 is wrong, and often only in its spelling or ending, for example bhí instead of bhíomar.

Casual conversation, strong dialect, old archive tapes

Much rougher: up to 1 word in 2 can be wrong. Treat these transcripts as a best-effort draft.

Choose Irish before you upload
Gaston can detect Irish automatically, but choosing Irish yourself is more reliable, especially for short recordings or ones that mix Irish and English.
English passages are transcribed too
The model is trained on Irish. English stretches inside Irish audio usually come out as text as well, but less accurately.
04 · Training

What went into it

We started from OpenAI Whisper large-v3 and fine-tuned the whole model on Irish speech: parliament committees, dialect recordings, read sentences, podcasts, radio and YouTube.

1.55 B
parameters, based on OpenAI Whisper large-v3
~1,890 h
of Irish training audio
~210 h
transcribed by people
~1,460 h
transcribed in-house, then filtered and brought to one consistent spelling style
~210 h
transcribed by earlier versions, kept only where an independent second system agreed
9.8 h
of music and noise, so the model writes nothing when nobody speaks
Speech in training
Parliament committees Dáil debates Dialect recordings Podcasts Radio YouTube Read speech Conversation

Audio from the test sets was kept out of every training version.

Trained for messy audio
Background noise Music Room echo Phone-line quality Speed changes Volume changes

Training clips are randomly altered this way, so the model copes with real uploads, not only studio audio.

05 · In Gaston

What you get

The model runs inside the Gaston transcription service, so every Irish recording becomes a transcript you can search, translate, bookmark and export.

á é í ó ú
Written Irish, not raw speech
Transcripts come with síntí fada and punctuation, and fillers and dialect spellings are evened out, so they read like written Irish.
Every word timestamped
Word-level alignment lets you click any word in the transcript and jump straight to that moment.
Recognises Irish
Our main language detector does not know Irish, so Gaston runs a second check that does. Choosing Irish yourself is still the safest option.
A 文
Translate it
Translate Irish transcripts into other languages right in the player. Translation from Irish is in beta.
Silence stays silent
Trained on music and noise with empty transcripts, so jingles and pauses do not turn into invented text.
Known weak spots
  • •Spontaneous conversation and strong dialect
  • •Old archive recordings
  • •English passages inside Irish audio
  • •Word endings and initial mutations
06 · Method

How we measured

  • Decoded the way the production service runs it, with beam search 5, and scored on normalised text.
  • The six test sets were frozen in October 2026 and kept out of every training version.
  • The public Common Voice training split contains sentences from its own test split. We removed 5,373 of them from training, so that score is not inflated.
Test set Kind of audio
Common Voice ga-IE
Mozilla · CC0
Short read sentences
FLEURS ga_ie
Google · CC-BY-4.0
Read Wikipedia sentences, many speakers
Oireachtas
Houses of the Oireachtas
Parliament committees, checked against the official record
Canúint
canuint.ie
Dialect archive, older recordings
abairt.com
abairt.com
Spontaneous conversation, Mayo dialect
YouTube
Gaston
Vlogs, interviews, multi-speaker video, scored against partly condensed subtitles

Credits

This model builds on openly shared Irish speech data and tools. Thank you to everyone who built and shared them.

  • Based on OpenAI Whisper large-v3 (MIT)
  • Houses of the Oireachtas (official record of proceedings)
  • Google FLEURS, CC-BY-4.0
  • Mozilla Common Voice, CC0
  • jimregan/wav2vec2-large-xlsr-irish-basic (word timestamps)
  • MUSAN corpus (background noise)
Dia duit

Try it on your own Irish recordings

Upload a file or paste a YouTube link, and Gaston does the rest.

Create an account