Hear a voice. Name the language. Six guesses.
Earshot is a daily language-guessing game. You hear a real recording of someone speaking and have six guesses to name the language. The first guess hears only the opening second and a half. Every wrong guess, or a skip, unlocks a longer stretch of the same recording, until the last guess hears all of it. A guess in the right language family turns amber, so even a miss tells you where to look. Everyone gets the same voice each day, and Unlimited mode plays clip after clip with a running score.
Press play and you hear a second and a half of a woman’s voice: soft, with a lot of “sh” sounds and vowels that seem to disappear into the nose. It sounds Romance, maybe. A reasonable first guess is French. It comes back amber, which means right family (Romance), wrong language. The clip now runs to three seconds. The nasal vowels are still there, but the rhythm is looser than French and the ends of words are swallowed. You try Catalan. Amber again, still Romance, but not it.
At five seconds you catch a word ending in a nasal “-ão” and a run of “-ções”, which only one Romance language really does. You guess Portuguese, and it goes green. Your result shows three guesses and how much you needed to hear, 5 seconds. The share grid shows only the colours, so it gives nothing away.
Green is the right language. Amber is a language from the same family as the answer: Spanish and Italian are both Romance, Polish and Czech are both Slavic, Finnish and Hungarian are both Uralic, and Amharic and Arabic are both Semitic. Red is a different family entirely. A skip is shown in grey and costs a guess, but it buys more audio without spending a guess on a language you do not believe in.
The families are deliberately narrower than the big textbook groupings. Indo-European, for example, is split into Romance, Germanic, Slavic, Celtic, Indo-Aryan, Iranian and others, so an amber result really does narrow things down. In all, the 159 languages you can guess from are sorted into 36 families.
Listen for sounds before words. Clicks point to southern and eastern Africa. A dense run of consonants suggests the Caucasus or Slavic languages. Tones, where the pitch of a syllable changes its meaning, are the mark of Mandarin, Cantonese, Vietnamese, Thai and many West African languages. Vowel harmony and long agglutinated words suggest Turkish, Hungarian or Finnish.
Rhythm helps too. Some languages give each syllable roughly the same length, as Spanish, Japanese and Yoruba do. Others squash unstressed syllables between stressed ones, as English, Russian and Dutch do. Even when you cannot place a single word, telling these patterns apart is often enough to get into the right family by your second guess.
The bank has 83 recordings in 81 languages, from widely spoken ones such as English, Mandarin, Spanish and Arabic to languages you may never have heard, like Lingala, Wolof, Cebuano, Twi and Luxembourgish. Every clip is a short excerpt of a real person speaking naturally, not reading a word list. They are trimmed, loudness-matched and re-encoded so that every clip plays at the same volume.
The recordings come from two sources. Wikitongues is a non-profit that records speakers of every language, shared under Creative Commons licences. FLEURS is a multilingual speech dataset published by Google researchers under CC BY 4.0. Each clip is credited with its source and licence on the page. The daily never plays the same language two days running, and Unlimited never plays that day’s daily clip.
Unlimited plays one clip after another, each with the same six-stage reveal. Naming a language on the first stage scores 6 points, on the second 5, and so on down to 1 point on the last. A miss scores nothing. Your session score, the number of rounds and your solved-in-a-row streak are shown at the top, and you can start a fresh session whenever you like.