Our speech model

Arc Audio

Arc Transcriber Space is the on-device ear behind Outpost Luna. Its main benchmark is real NASA speech: Artemis crew audio and NASA podcasts, with astronauts, flight controllers and scientists talking about the Moon. Below: the data, how test sets stay clean, and the scores.

Uses NASA open media; not endorsed by NASA. Model weights CC BY-NC 4.0 (see The model, below).

Why this corpus

General speech models stumble on exactly the words Outpost Luna needs most.

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.

Sources. 16 kHz mono audio in short segments (most 1.5–4 s), each with a lower-case spoken-form transcript.
SourceFromWhat it isEpisodes / recordingsHoursSegments
Artemis audioNASA Artemis audio libraryArtemis II crew and mission audio984.55,811
Houston We Have a PodcastNASA Johnson Space CenterInterviews with astronauts, flight controllers and engineers446356.6319,599
Small Steps, Giant LeapsNASA APPELConversations about mission lessons and project work18470.163,284
Curious UniverseNASAStories about NASA science and exploration10938.034,282
Total837≈469422,976

Hours by source

Train and test data

What trains the model and what tests it. Test material was set aside first.

Our NASA speech corpus. Hours of audio, episodes or recordings in brackets. This model was not trained on any of it.
PartTest (held out)Training split (next model)Removed as possible leaks
Houston We Have a Podcast10.3 h (11)328.8 h (415)17.5 h (20)
Small Steps, Giant Leaps1.7 h (5)66.9 h (177)1.5 h (2)
Curious Universe0.9 h (3)36.6 h (105)0.5 h (1)
Artemis audio15 recordings, 1,141 utterances*3.7 h (89)0.7 h (9)
Total12.8 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.

The other sources
SourceWhat it isIn trainingTestHow overlap is prevented
Common Voice (English)Everyday read speech, many speakersno2,000 clips, 3.3 hTest only: a fixed sample, never trained on.
Open read speechOpenly licensed read speech (Apache-2.0)yes (train split)300 utterancesThe corpus's own split: trained on its train split only.
Space questions (synthetic)Generated questions about the Moon, with app playback in the backgroundyes300 utterancesHeld-out questions, never trained on.
Artemis II crew audioCrew lunar-observation audioyes (other recordings)1,141 utterancesSplit by whole recording; the 15 test recordings never train.
Apollo mission audioApollo air-to-ground and mission audioyes500 utterancesHeld-out utterances, never trained on.
NASA TV and older podcast audioCommentary and interviewsyes–Training only.
Other open speechOpenly licensed general speech corporayes–Training only.
CREMA-DActed adult speechno600 clipsTest 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.

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.

Word error rate per test set
Test setSizeNVIDIA baseArc Transcriber Space
NASA speech corpus, held out
Houston We Have a Podcast11 episodes10.0%7.2%
Small Steps, Giant Leaps5 episodes9.7%6.4%
Curious Universe3 episodes6.0%4.8%
All podcast parts19 episodes, 12.8 h9.7%7.0%
Artemis audio1,141 utterances28.5%12.6%
Other test sets
Apollo mission audio50035.5%15.7%
Open read speech30042.0%18.5%
Space questions (synthetic)30048.1%9.1%
Common Voice 17 English2,00012.6%17.1%
CREMA-D adults6003.7%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).

Names metrics, Arc Transcriber Space
Test setName recallLinked correctlyLinked to a wrong entrySpace-tag F1NVIDIA base
Artemis II crew audio (1,141)93.4%94.5%1.0%0.88n/a
Space questions, synthetic (300)82.8%91.4%0.0%0.83n/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.

Arc tags

On top of the words: inline tags on names and numbers, utterance and moment tags, audio events, and a flag for speech meant for the app. 249 tags in 9 groups; Outpost Luna uses a few.

The 9 tag groups
GroupWhat it coversTags
MetaSignal properties: is someone speaking, overlapping voices, degraded or synthetic audio, language.7
SpeakerWho is talking: speaker roles and speaker changes.18
DiscourseWhat the speech does: statement, question, request, answer, and whether it is addressed to the device.73
ContentWhat kind of content it is: live action, studio talk, interview, ads, and its topic.66
EventsThings that just happened or are about to: sports plays, courtroom rulings, meeting decisions.23
SceneBackground sound: cheering, applause, music, laughter, whistles, buzzers, silence.12
AffectHow it sounds: emotion and arousal.8
SafetyToxic language and profanity, for filtering.16
EntitiesNamed things and numbers: people, organisations, places, products, dates, counts, measures.26
Total249

The tags Outpost Luna relies on

Precision / recall / F1 on the app's path, full test sets; gold tag count in brackets.

Inline tags: meaning and test scores (P / R / F1)
TagMeaningArtemis II crew audio (1,141)Space questions (300)
ENTITY_LUNARA named feature or site on the Moon: crater, mare, mountain, rille, landing site.88 / .88 / .88 (584).88 / .81 / .84 (124)
ENTITY_SPACECRAFTA named spacecraft, vehicle, rocket or module.86 / 1 / .92 (6).83 / .83 / .83 (29)
ENTITY_MISSIONA named space mission or programme, e.g. Apollo 11, Artemis II–.81 / .79 / .80 (43)
ENTITY_CELESTIALA named celestial body: the Moon, a planet, star or comet.92 / .97 / .94 (34)1 / 1 / 1 (2)
ENTITY_TERMA space science or engineering term.82 / .83 / .82 (115)1 / 1 / 1 (2)
ENTITY_PERSONA person named: crew, scientist or other.22 / .94 / .36 (47).13 / 1 / .22 (1)
NUMBERA spoken number with no more specific kind.51 / .67 / .58 (162).88 / 1 / .93 (7)
NUMBER_PERCENTA percentage0 / 0 / 0 (5)1 / 1 / 1 (13)
NUMBER_TIMEA time of day.18 / .06 / .09 (36)–
NUMBER_MEASUREA physical measurement: distance, speed, mass, temperature0 / 0 / 0 (14)0 / 0 / 0 (1)
All space name tags.88 / .88 / .88 (624).86 / .81 / .83 (198)

Weak spots: ENTITY_PERSON fires too often (low precision), and number kinds (time, measure) are rarely right. Strong: the four space name tags (lunar, spacecraft, mission, celestial), the ones the app links to its names lists. Other inline tags (organisation, place) have no labels in the space data.

Utterance flag and audio events the app uses. Not scored: the space data has no hand labels for them.
OutputMeaning
device_directedThe utterance is addressed to the app, not to another person (yes when p ≥ 0.5)
scene.silenceIt is nearly silent
meta.degraded_audioThe audio sounds bad: thin phone-line sound, heavy noise or distortion
meta.channel.telephoneThe audio has passed through a telephone channel
meta.channel.narrowbandThe audio is band-limited (about 8 kHz sampling or less)
meta.clippingThe signal is clipped or saturated
All 249 tags, with their meanings

Meta 7

  • meta.speech Someone is talking right now.
  • meta.overlap Two people are talking over each other at the same time.
  • meta.degraded_audio The audio sounds bad: thin phone-line sound, heavy noise or distortion.
  • meta.synthetic_voice The voice is computer-generated text-to-speech.
  • meta.lang.en The speech is in English.
  • meta.lang.non_en The speech is in a foreign language such as Spanish, Hindi or French.
  • meta.lang.mixed The speaker switches between English and another language.

Speaker 18

  • speaker.role.play_by_play This speaker calls the live action as it happens.
  • speaker.role.analyst This speaker is the expert who explains and evaluates, not the one calling the plays.
  • speaker.role.host This speaker presents the show and hands off to others.
  • speaker.role.co_host This speaker is the second regular presenter alongside the host.
  • speaker.role.reporter This speaker is a journalist reporting from the field or sideline.
  • speaker.role.guest This speaker is an invited guest or interviewee on a news or talk show.
  • speaker.role.athlete This speaker is a player or competitor talking during a sports or esports broadcast.
  • speaker.role.coach_staff This speaker is a coach, manager or team staff member, not a player.
  • speaker.role.game_official This speaker is the referee, umpire or match official.
  • speaker.role.caller_audience This speaker is a member of the public: a caller, fan or eyewitness.
  • speaker.role.announcer_voiceover This voice is an announcer or voice-over, not a presenter: ads, promos, PA or narration.
  • speaker.role.participant This speaker is an ordinary meeting participant, not the chair.
  • speaker.role.judge This speaker is the judge or justice presiding.
  • speaker.role.counsel This speaker is a lawyer for one of the parties.
  • speaker.role.witness This speaker is a witness testifying under oath.
  • speaker.role.device_user This speaker is giving a command to a voice assistant.
  • speaker.role.other This speaker has none of the usual broadcast, meeting, court or assistant roles.
  • speaker.change A different person just started speaking.

Discourse 73

  • discourse.act.statement The speaker states a fact or tells what happened.
  • discourse.act.question The speaker asks a question.
  • discourse.act.opinion_assessment The speaker gives an opinion or judges something.
  • discourse.act.suggestion_proposal The speaker proposes what the group could do.
  • discourse.act.commitment_offer The speaker promises or offers to do something.
  • discourse.act.command_request The speaker tells someone to do something now.
  • discourse.act.ritual_procedural A conventional formula: greeting, thanks, goodbye, oath or courtroom formality.
  • discourse.answer The speaker is answering a question someone just asked.
  • discourse.backchannel A short listener response like mm-hmm, yeah or right.
  • discourse.interruption Someone cuts in and the current speaker stops.
  • discourse.turn_yield The speaker finishes their turn and hands the floor to someone else.
  • discourse.assistant.device_directed This speech is addressed to the voice assistant.
  • discourse.assistant.command_complete The voice command has just ended.
  • discourse.intent.alarm_query The user asks what alarms are set.
  • discourse.intent.alarm_remove The user asks to delete an alarm.
  • discourse.intent.alarm_set The user asks to set an alarm.
  • discourse.intent.audio_volume_down The user asks to turn the volume down.
  • discourse.intent.audio_volume_mute The user asks to mute the sound.
  • discourse.intent.audio_volume_other The user asks to change some other audio setting.
  • discourse.intent.audio_volume_up The user asks to turn the volume up.
  • discourse.intent.calendar_query The user asks what is on their calendar.
  • discourse.intent.calendar_remove The user asks to delete a calendar event.
  • discourse.intent.calendar_set The user asks to add an event or reminder to the calendar.
  • discourse.intent.cooking_query The user asks a cooking question.
  • discourse.intent.cooking_recipe The user asks for a recipe.
  • discourse.intent.datetime_convert The user asks to convert a time between time zones.
  • discourse.intent.datetime_query The user asks for the date or time.
  • discourse.intent.email_addcontact The user asks to add an email contact.
  • discourse.intent.email_query The user asks to check their emails.
  • discourse.intent.email_querycontact The user asks for a contact's details.
  • discourse.intent.email_sendemail The user asks to send an email.
  • discourse.intent.general_greet The user greets the assistant.
  • discourse.intent.general_joke The user asks the assistant to tell a joke.
  • discourse.intent.general_quirky The user says something playful or odd to the assistant.
  • discourse.intent.iot_cleaning The user asks the robot vacuum to clean.
  • discourse.intent.iot_coffee The user asks the coffee machine to make coffee.
  • discourse.intent.iot_hue_lightchange The user asks to change the light colour.
  • discourse.intent.iot_hue_lightdim The user asks to dim the lights.
  • discourse.intent.iot_hue_lightoff The user asks to switch the lights off.
  • discourse.intent.iot_hue_lighton The user asks to switch the lights on.
  • discourse.intent.iot_hue_lightup The user asks to make the lights brighter.
  • discourse.intent.iot_wemo_off The user asks to switch a smart plug off.
  • discourse.intent.iot_wemo_on The user asks to switch a smart plug on.
  • discourse.intent.lists_createoradd The user asks to create a list or add an item to it.
  • discourse.intent.lists_query The user asks what is on a list.
  • discourse.intent.lists_remove The user asks to remove an item or list.
  • discourse.intent.music_dislikeness The user says they dislike the current song.
  • discourse.intent.music_likeness The user says they like a song or artist.
  • discourse.intent.music_query The user asks what song is playing.
  • discourse.intent.music_settings The user asks to change music playback settings like shuffle or repeat.
  • discourse.intent.news_query The user asks for the news.
  • discourse.intent.play_audiobook The user asks to play an audiobook.
  • discourse.intent.play_game The user asks to play a game.
  • discourse.intent.play_music The user asks to play music.
  • discourse.intent.play_podcasts The user asks to play a podcast.
  • discourse.intent.play_radio The user asks to play a radio station.
  • discourse.intent.qa_currency The user asks an exchange rate question.
  • discourse.intent.qa_definition The user asks what a word means.
  • discourse.intent.qa_factoid The user asks a general knowledge question.
  • discourse.intent.qa_maths The user asks a maths question.
  • discourse.intent.qa_stock The user asks about a stock price.
  • discourse.intent.recommendation_events The user asks for events to go to.
  • discourse.intent.recommendation_locations The user asks for a place recommendation like a restaurant.
  • discourse.intent.recommendation_movies The user asks for a movie recommendation.
  • discourse.intent.social_post The user asks to post something on social media.
  • discourse.intent.social_query The user asks about social media updates.
  • discourse.intent.takeaway_order The user asks to order takeaway food.
  • discourse.intent.takeaway_query The user asks about a takeaway order or menu.
  • discourse.intent.transport_query The user asks about train or bus times.
  • discourse.intent.transport_taxi The user asks to book a taxi.
  • discourse.intent.transport_ticket The user asks to book a travel ticket.
  • discourse.intent.transport_traffic The user asks about the traffic.
  • discourse.intent.weather_query The user asks about the weather.

Content 66

  • content.mode.live_action The broadcast is covering live game play right now.
  • content.mode.replay_highlight A replay or highlight of earlier action is being shown.
  • content.mode.studio_talk Presenters and analysts talk in the studio, away from live action.
  • content.mode.field_report A reporter files a report from a location.
  • content.mode.interview The segment is an interview: questions and answers with one guest.
  • content.mode.call_in Listeners phone in to the show.
  • content.mode.live_event_feed The broadcast carries a live event feed such as a press conference or courtroom.
  • content.mode.ad_promo A commercial, sponsor message or promo is airing.
  • content.mode.bumper_jingle A short station jingle or show theme sting plays.
  • content.mode.music_program Music is being played as content, like a song on the radio.
  • content.mode.break_intermission The event is on a scheduled break or intermission with no programme talk.
  • content.mode.no_commentary The event continues but nobody is commentating.
  • content.game_state.in_play The ball or puck is live and play is running.
  • content.game_state.stoppage Play is briefly stopped between plays or after a whistle.
  • content.game_state.timeout A team or TV timeout is in effect.
  • content.game_state.review Officials are reviewing a play on video.
  • content.game_state.period_break It is the break between periods, quarters or halves.
  • content.game_state.pre_post The game has not started yet or is already over.
  • content.legal_phase.argument_petitioner The petitioner's lawyer is arguing before the appellate court.
  • content.legal_phase.argument_respondent The respondent's lawyer is arguing before the appellate court.
  • content.legal_phase.rebuttal The petitioner's lawyer is giving rebuttal at the end.
  • content.legal_phase.opening_statement A lawyer gives an opening statement to the jury.
  • content.legal_phase.direct_exam A lawyer questions their own witness on direct examination.
  • content.legal_phase.cross_exam The opposing lawyer cross-examines the witness.
  • content.legal_phase.redirect_recross Redirect or recross examination after cross.
  • content.legal_phase.closing_argument A lawyer gives the closing argument to the jury.
  • content.legal_phase.bench_procedure The judge handles procedure: rulings, sidebars, instructions, scheduling or the verdict.
  • content.meeting_phase.opening The meeting is opening with greetings and the agenda.
  • content.meeting_phase.agenda_admin The meeting handles admin items: roles, budget, requirements, process.
  • content.meeting_phase.presentation One participant presents prepared material.
  • content.meeting_phase.discussion The group openly discusses and evaluates options.
  • content.meeting_phase.chitchat Off-topic social chat in the meeting.
  • content.meeting_phase.closing The meeting is wrapping up with next steps and goodbyes.
  • content.sports_focus.game_action The commentary describes the play happening now.
  • content.sports_focus.analysis_tactics The commentary explains tactics and why a play worked.
  • content.sports_focus.player_team_background The commentary tells background stories about players or teams.
  • content.sports_focus.stats_records_standings The commentary cites statistics, records or standings.
  • content.sports_focus.injury_health The commentary is about an injury or a player's health.
  • content.sports_focus.officiating_rules The commentary discusses referees' calls or the rules.
  • content.sports_focus.off_game_banter The commentators chat about something unrelated to the game or sport.
  • content.subject.politics_government The story is about politics or government.
  • content.subject.economy_business The story is about the economy or business.
  • content.subject.crime_justice The story is about crime, police or the courts.
  • content.subject.conflict_security The story is about war, terrorism or the military.
  • content.subject.disaster_accident The story is about a disaster or accident.
  • content.subject.weather_climate The story is about the weather or climate.
  • content.subject.science_tech The story is about science or technology.
  • content.subject.health_medicine The story is about health or medicine.
  • content.subject.education_society The story is about education or social issues.
  • content.subject.sports The story is about sports.
  • content.subject.arts_entertainment The story is about entertainment or culture.
  • content.subject.lifestyle_personal The talk is about personal life or lifestyle.
  • content.story_start A new story or topic begins here.
  • content.breaking_news The presenter announces breaking news.
  • item.esports_title.league_of_legends The esports broadcast is League of Legends.
  • item.esports_title.valorant The esports broadcast is Valorant.
  • item.esports_title.counter_strike The esports broadcast is Counter-Strike.
  • item.esports_title.dota2 The esports broadcast is Dota 2.
  • item.esports_title.rocket_league The esports broadcast is Rocket League.
  • item.esports_title.call_of_duty The esports broadcast is Call of Duty.
  • item.esports_title.other_title The esports broadcast is a game other than the main listed titles.
  • item.case_area.criminal The case is about criminal law.
  • item.case_area.civil_rights_liberties The case is about civil rights or civil liberties.
  • item.case_area.economic_regulatory The case is about business, labour, tax or regulation.
  • item.case_area.government_powers The case is about powers of government and courts.
  • item.case_area.private_civil The case is a private civil dispute.

Events 23

  • event.sports.score.recent A team scores - it just happened seconds ago.
  • event.sports.score.imminent Before it happens: a scoring chance is building up; a team scores may follow within half a minute.
  • event.sports.touchdown.recent An NFL touchdown is scored - it just happened seconds ago.
  • event.sports.touchdown.imminent Before it happens: the offense is driving deep toward the end zone; an NFL touchdown is scored may follow within half a minute.
  • event.sports.field_goal.recent An NFL field goal is made - it just happened seconds ago.
  • event.sports.field_goal.imminent Before it happens: the kicking unit lines up for an attempt; an NFL field goal is made may follow within half a minute.
  • event.sports.wicket.recent A cricket batter is out - it just happened seconds ago.
  • event.sports.wicket.imminent Before it happens: the bowler is building pressure on the batter; a cricket batter is out may follow within half a minute.
  • event.sports.boundary.recent A cricket boundary: four or six runs - it just happened seconds ago.
  • event.sports.boundary.imminent Before it happens: the batter is shaping to attack the bowling; a cricket boundary: four or six runs may follow within half a minute.
  • event.sports.turnover.recent Possession is lost to the other team - it just happened seconds ago.
  • event.sports.penalty_foul.recent The official calls a penalty or foul - it just happened seconds ago.
  • event.sports.timeout_called.recent A timeout is called - it just happened seconds ago.
  • event.sports.review_challenge.recent Officials start a video review or challenge - it just happened seconds ago.
  • event.sports.injury_stoppage.recent Play stops because a player is injured - it just happened seconds ago.
  • event.sports.substitution.recent A soccer substitution is made - it just happened seconds ago.
  • event.sports.period_end.recent A period, quarter, half or innings ends - it just happened seconds ago.
  • event.sports.shot_on_target.recent A soccer shot on target is saved or blocked - it just happened seconds ago.
  • event.legal.objection.recent A lawyer stands and objects to a question or evidence - it just happened seconds ago.
  • event.legal.ruling_sustained.recent The judge agrees with the lawyer and blocks the question - it just happened seconds ago.
  • event.legal.ruling_overruled.recent The judge rejects the lawyer's challenge and lets the question stand - it just happened seconds ago.
  • event.meeting.decision.recent The meeting reaches a decision - it just happened seconds ago.
  • event.meeting.action_item.recent A task is assigned to someone in the meeting - it just happened seconds ago.

Scene 12

  • scene.cheer The crowd cheers and roars.
  • scene.applause People clap their hands.
  • scene.crowd_ambience Background crowd murmur without cheering.
  • scene.boo_jeer The crowd boos.
  • scene.chant The crowd chants in rhythm.
  • scene.music Music is playing.
  • scene.laughter People laugh.
  • scene.horn A horn sounds, like a goal horn or air horn.
  • scene.whistle A whistle blows.
  • scene.buzzer A buzzer sounds.
  • scene.gunshot_explosion Gunfire or an explosion is heard.
  • scene.silence It is nearly silent.

Affect 8

  • affect.emotion.neutral The voice sounds emotionally neutral.
  • affect.emotion.happy The voice sounds happy or joyful.
  • affect.emotion.angry The voice sounds angry.
  • affect.emotion.sad The voice sounds sad.
  • affect.emotion.surprised The voice sounds surprised.
  • affect.arousal.low The speaker sounds calm and low-energy.
  • affect.arousal.medium The speaker sounds engaged at a normal energy.
  • affect.arousal.high The speaker sounds very excited or agitated.

Safety 16

  • safety.toxicity.insult_harassment The speaker insults or harasses someone.
  • safety.toxicity.hate_identity The speaker attacks a group based on identity.
  • safety.toxicity.threat_violence The speaker threatens violence.
  • safety.toxicity.sexual_explicit The speaker says something sexually explicit.
  • safety.toxicity.self_harm The speaker encourages self-harm or suicide.
  • safety.toxicity.toxic_any Any toxic speech.
  • safety.profanity.mild A mild swear word like damn or hell.
  • safety.profanity.strong A strong swear word or obscenity.
  • safety.pii.phone_number A private person says a phone number.
  • safety.pii.email A private email address is spoken.
  • safety.pii.street_address A private street address is spoken.
  • safety.pii.private_name The name of a private, non-public person is spoken.
  • safety.pii.card_account_number A card or account number is spoken.
  • safety.pii.gov_id A government ID number is spoken.
  • safety.pii.url_handle A private social media handle or URL is spoken.
  • safety.pii.date_of_birth A private person's date of birth is spoken.

Entities 26

  • entity.mention.person A person is mentioned by name.
  • entity.mention.team A sports or esports team is named.
  • entity.mention.organization An organisation is named: company, league, agency, court or party.
  • entity.mention.place A place is named: country, city, venue or street.
  • entity.mention.work_product A product, brand, app or title of a work is named.
  • entity.mention.event_name A named event is mentioned.
  • entity.mention.legal_ref A legal reference: case name, statute or constitutional clause.
  • entity.mention.game_entity A video-game character, item or objective is named.
  • entity.mention.game_map A video-game map is named.
  • entity.numeral A number is being spoken.
  • entity.number.score A game score is spoken.
  • entity.number.clock The game clock is spoken.
  • entity.number.time_of_day A time of day is spoken.
  • entity.number.date A date or year is spoken.
  • entity.number.duration A length of time is spoken.
  • entity.number.money An amount of money is spoken.
  • entity.number.percent A percentage is spoken.
  • entity.number.ordinal An ordinal or rank is spoken.
  • entity.number.count A count of things is spoken.
  • entity.number.stat A player or team statistic is spoken.
  • entity.number.measure A physical measurement is spoken.
  • entity.number.jersey_number A jersey number is spoken.
  • entity.number.field_position A field position or yard line is spoken.
  • entity.number.ratio_odds A ratio or betting odds are spoken.
  • entity.number.ref_number A reference number is spoken: page, exhibit, statute section, channel.
  • entity.number.id_pii A private identifying number is spoken.

Use it

Model Hugging Face

Weights, model card and sample clips.

WhissleAI/arc-transcriber-space

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.md

Licence 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.