How the index was made

What on this site is the archive's catalogue, what is the project's own work, and what was done by a machine — so that a reader knows how far to trust each thing.

The catalogue

Every tape on Media Collections Online carries an introduction written by the Archives of Traditional Music — date, place, who is interviewed and by whom, part x of y — and a series of timed descriptions, 8,019 in all, each saying in a sentence or two what happens in the next few minutes. These descriptions are the backbone of the site: the passages on every tape page are the catalogue's, word for word, and the time links open the tape where the catalogue says the passage begins. Three quarters of the tapes are described in sentences; for a quarter (Zhytomyr, Kyiv, Ovruch, Slavuta, Chernihiv and a few others) the descriptions are one-word labels and the substance is in the tape introduction, which the tape page shows beside the player.

The catalogue is not free of error. Dates are placeholders on one group of tapes (the Bershad' record carries dates in 1903 and 1904), some interviewees are not named in the introduction of a tape they appear on, names are spelled in several ways, and the times of a few passages run past the end of the tape. Where the project has corrected something, the correction is applied and noted; the rest is shown as the catalogue has it, with a mark where it is known to be unreliable.

People and places

The catalogue names people as strings — 580 different ones for roughly 380 people. The project merged them into persons by surname, given name and patronymic within a record, and across records where the catalogue's own cross-references ("Part 2 of 3, see MDV 67") link the tapes; spouses and relatives who share a surname were kept apart. A handful of merges are marked as probable rather than certain, and the person's page says so. Towns were normalized to one spelling each, given coordinates, and — where the town is also a point of the Language and Culture Atlas of Ashkenazic Jewry — linked to the atlas's informant location. Yiddish names of towns follow Paul (Hershl) Glasser's Yiddishland gazetteer (YIVO Institute for Jewish Research); where he gives several forms, the first is shown and the others follow in parentheses, with his transcription of the pronunciation. A few towns he does not list carry a name marked unverified, and corrections are welcome. Beside the English name every town carries its name in the language of the country it lies in today — Ukrainian, Romanian, Hungarian, Slovak — and, in smaller type, the other names a reader may know it by: the Russian name the speakers themselves used in Soviet documents, the Polish, Hungarian or Romanian name where the town has one, the Hebrew name, and former names; all of these come from OpenStreetMap (© OpenStreetMap contributors), and the search finds a town under any of them. The site's own language is English for now; the names are the first layer of a gradual opening to readers of Ukrainian, Russian, Hebrew, Hungarian and Romanian. The map draws the country outlines from Natural Earth and marks every town where the project recorded, in 2002–2012 or 2012–2018, with a dot sized by the hours recorded there; a town of the later recordings is drawn only once the project's index names it securely enough to be given its official name (Varshe is Warsaw, Kroke is Kraków, Ger is Góra Kalwaria), and one still being identified is listed but not drawn.

Conversations below the tape

A tape is the unit of the catalogue and of the online copies, but not of the fieldwork: the team sometimes recorded two interviews on one tape, or drove from one town to the next with the camera running, and an interview often ran over two or three tapes. So the site also works with a second unit, the conversation — one speaker or group, one place, one day — defined as a run of the catalogue's own passages on a tape, continuing over the following tapes where the catalogue's part numbers or the tape numbers say so. Nothing is cut: a town page lists the conversations recorded there with their hours added up from the actual stretches, a speaker's page lists their conversations across tapes, and a tape page says where one conversation ends and the next begins. Where a tape holds one conversation in one town, which is the case for 574 of the 746 tapes' parts, nothing had to be inferred. Where it holds more, the division was drawn from the catalogue's descriptions — the interviewees' surnames in the passages ("Kalussky talks…", then "the second part of the recording is a formal interview with Faiershtein") — and checked against the machine transcript, whose list of speakers gives time ranges; where both agree the part is marked two sources agree, where only the descriptions speak proposed, and where the signals conflict uncertain. On a tape the catalogue places in two towns, the passages' own place names decide which part belongs where when they can; otherwise the hours are split between the towns and marked approximate. All of this can be corrected, and the open questions page lists what is still open.

Where the speakers were born

The catalogue names a birth place for 266 of the 393 speakers, usually in the interviewer's transliteration ("village of Sguritzy", "Chipovich (Zhytomyr region)"). Where the string names a town of the recordings it was recognised at once; the others were looked up by name in OpenStreetMap's gazetteer, within the region and, where the catalogue gives a hint, near the town it names, and checked against Glasser's Yiddishland and against what the machine transcript's speaker list says about the speaker's birth. A place found by its name and confirmed by the transcript is shown without comment; one found by name alone is marked proposed; one whose modern name had to be guessed from the catalogue's spelling is marked guessed — to confirm and drawn hollow on the map; a string that names many places, or none that could be found, is left unplaced and listed among the open questions. 247 of the 266 speakers with a birth place are on the second map, where the speakers were born; 146 of them were born somewhere other than where they were recorded. The birth place is what the catalogue gives; the place where a speaker grew up and acquired their Yiddish, which matters more for the dialect questions, is not yet recorded separately.

Topics

The 51 topics are a vocabulary derived from the descriptions themselves: the most frequent subjects were counted, grouped and glossed, and three language models then read every description independently and assigned topics from that fixed list. A topic was kept when at least two of the three agreed; a topic only one model gave was set aside. A topic therefore means "the description mentions this" — not "the speaker says this", and not "this is absent elsewhere". Concrete subjects (holidays, songs, Passover, schooling, food) were assigned very consistently; general ones (daily life, the Soviet system, identity) less so, and their counts should be read as indications.

Songs

The song index is the project's own: 0 songs, recitations and melodies noted during the fieldwork with the tape and the minute at which each begins. The titles are the opening words as the index-maker heard them, which is why one song appears under several spellings. Each entry was re-linked to the digitized tapes and checked against the catalogue's descriptions (797 of the 851 timed entries fall in a passage the catalogue describes as singing), and each was cut as a clip for the project's YouTube channel; the clip begins a few seconds before the noted time and ends at the next song, the next described passage or after four minutes, whichever comes first — so the end of a clip is a guess, and the full tape is always a click away.

The questionnaire

The catalogue marks the stretches in which the dialect questionnaire is put but not which items were asked. In October 2026 the project ran a speech recognizer adapted to AHEYM's recordings over those stretches and had a language model read the machine transcripts for the questions: the Russian prompt, what it asks for, the kind of question, and the Yiddish form heard in reply. The questions were grouped into items and, where possible, keyed to the corresponding item of the LCAAJ questionnaire. What came out is reliable about when and of whom a question was asked and mostly reliable about what was asked; the forms heard in reply are the recognizer's reading, with a word error rate that makes them unfit to quote, and they are not shown until they have been checked by ear. The questionnaire pages are a finding aid to the moments, not a dialect atlas — the atlas is what a listener can now make from them.

The transcript layer

In October 2026 the project ran a speech recognizer adapted to AHEYM's recordings over the whole MiniDV collection — a Yiddish pass with word times, a multilingual pass for the Russian, Hungarian, Romanian, Polish and English stretches, and speaker diarization, with the interviewers found by voice — and had a language model read the result, tape by tape, grouped under the catalogue's own passages. For 7,898 of the 8,019 passages it wrote a description of what the transcript shows the passage to be about, in the catalogue's style, with topics from the same list, the persons and places heard, and a note of what the transcript adds to the catalogue's entry; for every tape a summary, the speakers as the transcript identifies them, the languages with their minutes, and notes on where catalogue and transcript disagree. The catalogue's entries, ids and times were not touched. On this site the layer sits under the catalogue's text, marked "from the machine transcript", and can be switched off on any tape page; where the catalogue has only a label — about a thousand passages on the Zhytomyr, Kyiv, Ovruch, Slavuta, Chernihiv and a few other records — it is the only description there is.

What it is worth: everything in it is a paraphrase of a machine reading. The recognizer gets a quarter to a third of the content words wrong on a clear indoor speaker and more than half on a hard outdoor one; names, numbers and family relations are its weakest points, and a "(?)" marks a doubt the model noticed, not every doubt there is. In the listening check of 25 October 2026, forty passages were heard against their descriptions: 31 held whole, 9 in part, none was wrong; the partly cases were a detail — a word glossed by its surface look, a relation confused, a year off by one, a topic from the next passage's first seconds — never the subject. The records described only by labels held at least as well as the others. So the topic of a passage can be trusted; a name, a number, a relation or a translated word is to be heard before it is cited. A description says what a passage is about, not what was said; nothing in it is a quotation. Counts drawn from this layer — topics, periods, places — are counts of what the machine transcript yielded and the model read into it, not of the recordings. Where a tape's languages are given, "Russian" includes speech with Ukrainian features. The layer is refreshed in part whenever the transcripts improve; each description records the run that wrote it.

The later expeditions of 2012–2018 — recordings held by the project and not online — are described from their transcripts alone, with a separate and weaker result: in the last complete check of that layer, 21 of 41 descriptions held whole, 17 in part and 3 were wrong, the faults being omissions, a place resolved to the wrong town, and stretches of Hungarian or Russian speech taken for Yiddish. Those pages say so.

Where the videos play from

The player on this site streams the project's copies on archive.org directly — the file itself, through the browser's own video element — and falls back to the project's YouTube channel where archive.org has not yet prepared a tape, which it does in its own time after an upload. Archive.org shows no advertisements, sets no cookies, restricts no country, and lets the player seek to the second and run one part of a conversation into the next across tapes; its servers are slower than YouTube's and the picture is modest (the copies are 640 by 480, as the MiniDV tapes were). Where both copies exist, a switch under the player lets you choose, and the choice is remembered in your browser. Tapes on neither service link to Indiana University's Media Collections Online, where every tape streams. The listing of what is where is renewed at every build of the site; 302 of the 746 tapes play here at the moment.

What is not here yet

The words spoken on the tapes are not searchable: machine transcripts of the whole collection exist, but their accuracy is not yet good enough to publish as text, and a Holocaust testimony misquoted by a machine is worse than one that must be listened to. The plan is to make the transcripts searchable first — a hit showing the minute, to be verified by ear — and to publish transcripts as text, and as subtitles on the videos, only once they have been corrected or the recognizer has improved enough.

Counts

Counts on this site are counts of catalogue entries, not of what was said: a town with forty tapes is not necessarily a town with forty interviewees, a topic with two thousand passages is a topic the cataloguers named often. Cite them as such.