Why this corpus
General speech models stumble on exactly the words Outpost Luna needs most.
- Real mission speech. Radio loops with noise, cross-talk and clipped phrasing. Clean read speech has none of this.
- The Moon's vocabulary. Tycho, Shackleton, Sinus Medii, Orion, Artemis II. A general model writes "tie co" for Tycho.
- Experts in plain language. NASA podcast guests explain the Moon the way learners ask about it.
- Long, unscripted talk. Whole episodes, nobody reading a script: a hard, honest test.
What is in the corpus
Team Outpost Luna collected and cleaned the corpus: NASA public audio with matching transcripts. Today it is our main test set; its training split is reserved for the next model.
| Source | From | What it is | Episodes / recordings | Hours | Segments |
|---|---|---|---|---|---|
| Artemis audio | NASA Artemis audio library | Artemis II crew and mission audio | 98 | 4.5 | 5,811 |
| Houston We Have a Podcast | NASA Johnson Space Center | Interviews with astronauts, flight controllers and engineers | 446 | 356.6 | 319,599 |
| Small Steps, Giant Leaps | NASA APPEL | Conversations about mission lessons and project work | 184 | 70.1 | 63,284 |
| Curious Universe | NASA | Stories about NASA science and exploration | 109 | 38.0 | 34,282 |
| Total | 837 | ≈469 | 422,976 | ||
Hours by source
Train and test data
What trains the model and what tests it. Test material was set aside first.
| Part | Test (held out) | Training split (next model) | Removed as possible leaks |
|---|---|---|---|
| Houston We Have a Podcast | 10.3 h (11) | 328.8 h (415) | 17.5 h (20) |
| Small Steps, Giant Leaps | 1.7 h (5) | 66.9 h (177) | 1.5 h (2) |
| Curious Universe | 0.9 h (3) | 36.6 h (105) | 0.5 h (1) |
| Artemis audio | 15 recordings, 1,141 utterances* | 3.7 h (89) | 0.7 h (9) |
| Total | 12.8 h (19) + Artemis | ≈436 h (786) | 20.3 h (32) |
* 15 whole Artemis recordings, kept out of every training set. The podcast test set is 19 whole episodes, 10,410 segments.
| Source | What it is | In training | Test | How overlap is prevented |
|---|---|---|---|---|
| Common Voice (English) | Everyday read speech, many speakers | no | 2,000 clips, 3.3 h | Test only: a fixed sample, never trained on. |
| Open read speech | Openly licensed read speech (Apache-2.0) | yes (train split) | 300 utterances | The corpus's own split: trained on its train split only. |
| Space questions (synthetic) | Generated questions about the Moon, with app playback in the background | yes | 300 utterances | Held-out questions, never trained on. |
| Artemis II crew audio | Crew lunar-observation audio | yes (other recordings) | 1,141 utterances | Split by whole recording; the 15 test recordings never train. |
| Apollo mission audio | Apollo air-to-ground and mission audio | yes | 500 utterances | Held-out utterances, never trained on. |
| NASA TV and older podcast audio | Commentary and interviews | yes | – | Training only. |
| Other open speech | Openly licensed general speech corpora | yes | – | Training only. |
| CREMA-D | Acted adult speech | no | 600 clips | Test only, never trained on. |
How it is annotated
Every segment carries the labels the model returns at run time, ready for training the next model.
- Spoken-form transcripts
- Written as said: "apollo twelve", "nineteen sixty nine". The app's names lists use the same form, so heard and linked text line up.
- Entity tags
- Inline by type: lunar feature, site, celestial body, spacecraft, mission, crew, numbers, places, organisations.
- Names lists
- About 9 in 10 segments carry a names list like the app's, so a model learns to link each mention or say "not in the list".
- Context
- Earlier text and names from the same episode travel with each segment: context to use, not depend on.
- Sounds and channel
- Beeps, static, music beds and cross-talk are labelled, not dropped. Speakers stay anonymous.
The full rules for names lists are in the names-list guide.
Keeping test sets out
A score means something only if the model never saw the test audio.
- Test recordings never enter. The 15 Artemis test recordings were skipped when the corpus was loaded.
- Whole episodes, checked against every test set. Each episode is searched for every test sentence of 6+ words (Artemis, Apollo, read speech, space questions, held-out names, Common Voice); a segment equal to a 4+ word test sentence also counts. One hit removes the whole episode from the training split. 32 removed (20.3 h): 9 Artemis recordings that rebroadcast test audio, and 23 podcast episodes that matched only common phrases.
- A new held-out test set. About 2.5% of each podcast series (at least 2 episodes), picked by a fixed hash. Only episodes with under 5% 8-word overlap with any earlier training text qualify. 19 episodes, 12.8 h, 10,410 segments (Houston We Have a Podcast 11, Small Steps, Giant Leaps 5, Curious Universe 3), frozen before any training on the corpus.
- General speech as a guard. A fixed Common Voice English sample measures what space tuning costs everyday speech. Reported below.
Results for Arc Transcriber Space
Word error rate per test set, against the NVIDIA FastConformer it started from. Both streaming, same text normalisation. Lower is better.
| Test set | Size | NVIDIA base | Arc Transcriber Space |
|---|---|---|---|
| NASA speech corpus, held out | |||
| Houston We Have a Podcast | 11 episodes | 10.0% | 7.2% |
| Small Steps, Giant Leaps | 5 episodes | 9.7% | 6.4% |
| Curious Universe | 3 episodes | 6.0% | 4.8% |
| All podcast parts | 19 episodes, 12.8 h | 9.7% | 7.0% |
| Artemis audio | 1,141 utterances | 28.5% | 12.6% |
| Other test sets | |||
| Apollo mission audio | 500 | 35.5% | 15.7% |
| Open read speech | 300 | 42.0% | 18.5% |
| Space questions (synthetic) | 300 | 48.1% | 9.1% |
| Common Voice 17 English | 2,000 | 12.6% | 17.1% |
| CREMA-D adults | 600 | 3.7% | 1.9% |
Common Voice: the base model is better (12.6% vs 17.1%). Tuning for space speech costs some everyday dictation. Every other test set improves.
Names: hearing and linking them
The app's own path: int8 model on a CPU, full test sets. Each names list holds the utterance's names plus 20 similar-sounding decoys. On sets with no names, any link is false. The base model cannot link names (n/a).
| Test set | Name recall | Linked correctly | Linked to a wrong entry | Space-tag F1 | NVIDIA base |
|---|---|---|---|---|---|
| Artemis II crew audio (1,141) | 93.4% | 94.5% | 1.0% | 0.88 | n/a |
| Space questions, synthetic (300) | 82.8% | 91.4% | 0.0% | 0.83 | n/a |
| Open read speech (300) | 0.67 false links per 100 utterances (no names in this set) | n/a | |||
| Apollo mission audio (500) | 0.4 false links per 100 utterances (no gold names in this set) | n/a | |||
Test with real voices: the model is built to ignore the app's own spoken answers, so text-to-speech voices can be partly ignored.
Use it
Python and C GitHub
The arc_transcriber package (the space extra) and a C library over ONNX Runtime.
pip install "arc-transcriber[space] @ git+https://github.com/WhissleAI/arc_transcriber"WhissleAI/arc_transcriber
Names lists read first
How to write the lists the app sends: kinds, order, aliases and the normalization the model uses.
docs/NAMES_LISTS.mdLicence and credits
- NASA
- Audio and transcripts come from NASA's public media: the Artemis audio library and the podcasts Houston We Have a Podcast, Small Steps, Giant Leaps and Curious Universe. Credit: NASA. Not affiliated with or endorsed by NASA.
- The corpus
- Collected and cleaned by Team Outpost Luna.
- The model
- © 2026 Whissle Inc. Weights licensed CC BY-NC 4.0 with the Whissle terms, for non-commercial use. Commercial licences: ksingla@whissle.ai. It builds on NVIDIA's FastConformer (CC-BY-4.0) and Moonshine (MIT); see the model card's NOTICE. Not endorsed by NVIDIA.