Since the advent of audio recording technology and the push towards documenting endangered languages, there has been an immense effort to document, describe, and preserve signed and spoken languages worldwide. However, literature on language documentation and description has rarely addressed how researchers should go about deciding what to transcribe and what the process of transcription should look like (Himmelmann 2018). Yet, issues relating to the practice of transcription have been discussed at length in other subfields of linguistics, such as conversation analysis and linguistic anthropology, where there has been a call for greater reflection on how one goes about creating a transcript of spoken discourse (e.g., Bucholtz 2000; Mondada 2007; Ochs 1979).

By recording the speech/signs and cultural practices of an individual or a group, it is possible to preserve their voice, likeness, and activities for posterity. Yet even with all the benefits of modern technologies, recording spoken discourse strips away contextual information from the original speech event, including the physical space, gestures, and audience. Additionally, recording transforms the discourse into a mobile, portable object that is circulable and mutable. In other words, the evanescent and transient nature of sound is concretized—or as linguistic anthropologists say, entextualized—when it becomes a physical record. Through inscription discourse gains new contextual meanings whenever it is revisited, a process often referred to as recontextualization.

Although the literature on entextualization and recontextualization is substantial (see Bauman & Briggs 1990; Silverstein & Urban 1996, and references therein), the same cannot be said for the implications of transcription practices within language documentation. Himmelmann (2018: 34, fn.2), references only Crowley’s 2007 Field linguistics: A beginner’s guide, which assumes that the reader uses a cassette recorder.1 In the section on text transcription, Crowley (2007: 138) writes, “[y]ou will need to transcribe exactly what was said on tape, as well as its meaning”. It bears asking what Crowley meant by “exactly what was said”, and whether obtaining an objective transcript is truly achievable (see Bucholtz 2000, who argues powerfully that it is not).2

As speech/signs are entextualized and recontextualized, multiple interpretations naturally arise (Austin 2017; Dobrin & Schwartz 2021; Haviland 1996; Urban 1996; Weber 2021). As we describe in this paper, a single audio recording may engender distinct interpretations not only by different individuals but also the same person on separate occasions. This finding indicates that the choice of who one works with, and when and where that work occurs, etc., can shape each working transcript. Furthermore, because new insights may surface every time a text is revisited, we suggest that transcription ideally should progress iteratively to reflect the evolution of the analysis. Since transcripts play a crucial role both in linguistic research and in the development of community-oriented resources, we encourage language documentarians to turn a self-reflective gaze on their transcription process. We offer several recommendations for how to go about this in the conclusion of this paper, where we also discuss the implications of the transcription process for transcription and digital archiving practices, and their effects on researchers and linguists, who almost always rely on transcripts to produce language revitalization materials and to study language structure and use.

Our proposal that transcription is an iterative process is not novel. In an 1889 paper, Franz Boas comments that his transcription of Inuktitut improved over time as his ability to discern the sounds of the language increased. Working on the rhetorical structure of narratives, Virginia Hymes (1987:68, emphasis hers) explicitly remarks, “what is involved is a patient working back and forth between content and form […] until gradually you arrive at an analysis that seems best at that stage of your knowledge of the narrative tradition and the particular narrator”. These examples underscore how the evolution of the researcher’s perspectives can necessitate the revision and re-revision of proposed analyses (see also Evans and Sasse 2007; Mondada 2007; Seidel 2016).

The idea that a single recording is subject to more than one interpretation is also not new. Researchers and community collaborators alike make conscious and unconscious decisions while transcribing, and transcribers themselves may also filter the content of a recording. For example, Dobrin (2021) compares transcriptions of texts in Bukiyip, a language of Papua New Guinea, with the original audio and identifies several discrepancies. In particular, the transcriber, a native speaker of the language, made decisions to purge repetitions and ‘purify’ the recording by removing Tok Pisin words, retranslating those portions into Bukiyip (see also Chelliah 2001; Haviland 1996; Jung & Himmelmann 2011; Urban 1996; Weber 2021).3

Given the number of potential transcripts, which of them should be maintained for future users? Although Jung and Himmelmann (2011: 219) state that “[t]here is no straightforward and easy answer to this question”, we propose that transcripts of all interpretations should be maintained and archived so long as it is ethical to do so (see Evans and Sasse 2007; Innes 2010). By preserving all versions of transcripts, we maintain each individual’s distinct but equally valid interpretation of the same text, which may potentially provide insights that would otherwise be lost. The structure of this paper is as follows. In Section 2, we describe the data analyzed, the project for which the data were collected, and the transcription process. In Section 3, we discuss the types of transcripts we use, how they reflect the goals of our project, and the challenges we encountered during the transcription process. In Section 4, we explore several types of variation found in the transcripts, from variation in the interpretation of the narratives to variation in forms in Apsáalooke (Crow) across different speakers. In Section 5, we examine the implications of this variation for transcription and archival practices. Finally, Section 6 offers concluding remarks, along with a set of considerations and lessons learned for other field researchers engaged in transcribing spoken discourse.

The data analyzed in this paper derives from the first author’s field recordings, which are housed in the California Language Archive of the Survey of California and Other Indian Languages maintained by the University of California, Berkeley (Alden et al., n.d.). Copies of these recordings have also been shared with the Little Big Horn College, the tribal college located in Crow Agency, the headquarters of the Crow Indian Reservation in Montana. Since 2018, the first author has been working with community collaborators to record oral histories pertaining to places on and around the reservation, along with personal reminiscences about changes to the reservation and the way of life, the contemporary language situation, and aspects of Apsáalooke identity and culture. As the Apsáalooke language is currently endangered, our project aims to address this issue in three ways. The first goal is to preserve Apsáalooke history, culture, and social structure as it exists in the early 21st century. The second is to create an online interactive storytelling map using Esri’s ArcGIS StoryMap, featuring the recorded narratives and accompanying transcripts. This platform enables community members to freely access and engage with the materials, thereby supporting future language revitalization and reclamation efforts. Finally, this project aims to document connected speech for more advanced learners of the language, who may be interested in how it is used in informal contexts, and to inform future work on understudied aspects of Apsáalooke language and culture.

To clarify the contributions of each author in the article, we use the pronoun ‘I’ to refer to efforts conducted by Edwin Ko, who has spent many years collaborating with community members on the Crow Indian Reservation to conduct fieldwork on the Apsáalooke language. The use of the first-person pronoun is not intended to minimize the contributions of Jem Burch, an undergraduate Yale linguistics student who served as a research assistant on the project. In fact, Jem Burch played a crucial role in the research, actively participating in the data collection during the summer of 2024, and significantly contributing to the analysis, literature review, drafting, and revising of the paper. The pronoun ‘we’ is employed when referring to collaborative efforts undertaken by both authors, emphasizing our joint responsibility and collaboration in the study’s development.

The project to record oral narratives emerged from sustained dialogue with community members, during which I carefully listened to their concerns. Although there were no explicit expectations regarding my role in language revitalization, community members frequently voiced anxieties not only about the future of the language but also about broader transformations in social structure, cultural practices, and relationships to the land, of which one of the most visible consequences is widespread urban decay, characterized by abandoned or dilapidated buildings and increasing family fragmentation. In addition, the project complements ongoing efforts at the tribal college, including a National Endowment for the Humanities-funded initiative to document oral histories from the post-World War II era. Although no formal approval process was established, the project was met with positive feedback from all community members with whom I engaged. Building on relationships formed during my initial visit to the community in 2016, when I collaborated with the then-Head of Education of the Crow Tribe, I recruited volunteers willing to share their stories. Individual sessions were conducted primarily at the Little Big Horn College library and the public library in the nearby border town of Hardin, or, in the case of place-based oral histories, at the relevant sites.

During a typical recording session, we would begin by inviting the community collaborator to introduce themselves and then ask them a range of open-ended questions about their perspectives on the Apsáalooke language and culture. These questions addressed topics such as the current state of the language, generational differences in how Apsáalooke is spoken, and what it means to be Apsáalooke when speaking to future generations. For the place-based oral histories, collaborators were asked to recall personal experiences and memories directly associated with specific extant and historical sites. In each case, collaborators were encouraged to answer in both English and Apsáalooke, and all responses were recorded. Following some of the sessions, we produced a transcript of the recording.

To illustrate our transcription process, we describe our transcription of a narrative contributed by community collaborator Theresa Sends Part Home, a woman born in the late 1950s, who had fond memories growing up in Lodge Grass, one of the towns on the reservation. I completed a first pass of the transcriptions using the time-aligned annotation tool ELAN (Wittenburg et al. 2006). During the COVID-19 pandemic, I collaborated remotely with Felice Big Day, an Apsáalooke speaker born in the late 1970s with whom I had previously worked, to transcribe the narratives. This interpretation session was conducted over the phone while I was in California and Felice was in Montana. To transcribe the recording, the entire audio of Theresa’s narrative was played in Apsáalooke. Then, Felice repeated and translated each word. Whereas the recording of a single narrative typically took place in a single session, the transcription process would often span several sessions. After the recordings had been transcribed in ELAN, I heavily revised them before transferring them to LaTeX to be formatted into interlinear text. The transcriptions were written in the “standard” Apsáalooke orthography, which omits aspects of rapid, informal speech. Recordings of all interpretation sessions were deposited at the California Language Archive before they were shared with the Little Big Horn College.

When it was safer to visit the reservation, I collaborated with Felice in person to reinterpret the same stories provided by Theresa. This time, rather than editing the ELAN file, I transcribed each word in my field notebook alongside the corresponding translations. After studying the recordings and my field notes, I made corrections to the original LaTeX file. After repeatedly listening to both Theresa’s recorded oral narrative and Felice’s recorded interpretation, I resolved any remaining issues with Felice. Once the transcriptions and glosses were finalized, they were transferred back to ELAN to be archived.

While this transcription process is largely consistent with that of other language documentarians (see, e.g., Chelliah and de Reuse 2011: 431–442), it differs from the practices used in other disciplines. As conversation analysts Hepburn and Bolden (2013: 57) state, “Conversation analytic transcripts need to be detailed enough to facilitate the analyst’s quest to discover and describe orderly practices of social action in interaction”. In the field of ethnopoetics, scholars aim to “[arrange] the transcript of an event of oral narration so as to reflect or recuperate the true rhetorical architecture of denotational text” (Moore 2013: 14). In these disciplines, transcription plays a fundamental role in allowing researchers to identify patterns that serve the specific communicative purpose being studied.

In contrast, the purpose of language documentation, as Woodbury (2010: 159) writes, is the “creation, annotation, preservation and dissemination of transparent records of a language”. That said, transparency is easier to aim for than to achieve. Although field notebooks may preserve some version of a given text, various challenges, such as idiosyncratic notation or missing contextual information, can reduce their transparency and accessibility for future users (see, e.g., Goddard 1976; Austin 2017; Begay et al. 2021). In many cases, archives contain only a single digital transcript, omitting any record of the iterative revisions that shaped the final version. Furthermore, even archival recordings that feature the voice of only a single speaker, as many of our Apsáalooke recordings do, cannot be treated as entirely monologic. As numerous scholars have argued, interviews are co-constructed by all participants, including the researcher, and are shaped by cultural expectations and the dynamics of the interaction itself (Briggs 1986; Haviland 1996; De Fina and Perrino 2011; Dobrin 2012).4

At every stage of the transcription process, some information inevitably gets lost. Miscommunication may occur—a researcher might mistake a word that underwent syncope for another word, or misinterpret false starts and hesitation markers as part of the previous or following words (see Jung and Himmelmann 2011: 211). After consulting with speakers, the researcher may discover their errors and correct them, but the revision process remains hidden if only the final version of the transcript is archived. It is human nature to be unwilling to reveal past mistakes, but presenting the evolution of transcripts can be insightful for others. Furthermore, if a researcher working with multiple speakers receives varying interpretations, but publishes only the interpretation of a single speaker, it is important to include in the metadata an explanation for why that particular interpretation was selected.

Every stage of the transcription process builds upon a previous stage. Even before listening to a recording, each user will contribute their own prior knowledge, lived experiences, and biases—what Fairclough (1989) and Bell (2011) call the “members’ resources” and “pre-view”, respectively. Therefore, we advocate that language documentarians archive multiple versions of each transcript. We also suggest that they provide relevant metadata, such as their research agenda and biography, to contextualize their collection for future users (Austin 2017; Dobrin & Schwartz 2021; Nathan & Austin 2004; Innes 2010; Woodbury 2014). For full disclosure, we have not archived all earlier versions of our transcripts; earlier drafts produced solely by me were replaced with later interpretations offered by speakers. It was only when I conceived the idea that laid the groundwork for this paper that we began preserving these initial drafts.

In this section, we describe our approach to transcribing Apsáalooke texts (Alden et al., n.d.) based on our goals—to provide transcripts for future researchers and language-learners—and reflect on its drawbacks.5^,6 For this, we find Bucholtz’s (2000: 1461–1463) distinction between naturalized and denaturalized transcripts particularly useful. Naturalized transcripts closely adhere to the standard written conventions, whereas denaturalized transcripts preserve aspects of the signed or spoken language, which may include incorporating paralinguistic features of signed and spoken language, such as pauses and intonation, into the transcripts. In some sense, naturalized transcripts are highly edited versions of denaturalized ones. However, as noted by Bucholtz (2000), both are extremes on a continuum, and neither is optimal. We have chosen to represent our texts in these two styles. Transcripts presented in a naturalized style expose learners of Aspáalooke to a standard written form without variation, fast-speech phenomena, etc.; transcripts presented in a denaturalized style allow advanced learners to study informal speech and provide data for future analysis of understudied areas of Apsáalooke linguistics.

The naturalized transcripts are written in the standard Apsáalooke orthography, which maintains information about pitch accent. However, these transcripts are not totally naturalized in the same way Bucholtz originally presented them, as we also preserve discourse features, such as hesitation markers, in order to highlight that the speech was performed orally rather than read from a script. The denaturalized transcripts preserve some fast-speech phenomena and other discourse features that would otherwise be omitted, including repetitions, false starts, self-corrections, and hesitation particles. However, we exclude intonation, duration of pauses, audible breathing, laughter, and other features that might be important to transcripts produced for conversation analysis and ethnopoetics. Our decision to exclude these features is not born out of a belief that such discursive features are unimportant, but as Ochs (1979: 44) remarks, “transcription is a selective process” that reflects one’s own goals. Being more selective about what we transcribe also reduces barriers of access to the transcripts, as including these additional annotations would decrease readability, especially for those focused on language revitalization.

Excluding paralinguistic features thus highlights the project’s overall purpose, which is primarily concerned with the content of the oral history and its importance to the Apsáalooke community. However, this comes at the expense of showcasing the speaker’s voice, that is, the “freedom to have one’s voice heard, freedom to develop a voice worth hearing” (Hymes 2004: 64; see also Dobrin 2012, 2021; Webster 2013). The presence of the interviewer is also omitted, based on the assumption that users are primarily interested in the voices of community members. Yet by presenting these transcripts as monologic, we inevitably obscure aspects of the participant framework—namely, the configuration of roles, statuses, and alignments occupied by each participant, including the researcher, in relation to others and to the unfolding interaction at a particular moment in time (Goffman 1981).

However, by offering two distinct transcripts, we are able to both cater to multiple audiences and escape the difficulty of choosing between a transcript that is overly technical and one that is overly simplistic (Jung and Himmelmann 2011: 207). Below, we provide examples of naturalized and denaturalized transcription styles from our collection, using excerpts from narratives offered by Theresa Sends Part Home as a case study.

The following four case studies illustrate the variability that emerges from our transcription process. The first compares interpretations conducted remotely and in-person with the same community collaborator, Felice Big Day. The second highlights two distinct interpretations of the same text also by Felice, received just five days apart. The third examines how two different collaborators produced divergent interpretations of the same recording. The final case focuses on interpretations of the same text from two different community collaborators, one of whom was the original contributor to that text. It should be noted that the differences observed in our transcripts, including those presented in this paper, primarily concern discourse markers or variants, rather than elements that would alter the core semantic content of the utterance, which is consistent with what has been observed by previous scholars (e.g. Jung and Himmelmann 2011; Webster 2013; see also Kowal and O’Connell 2014). In what follows, all transcripts are presented in a denaturalized style.

Given the variability in interpretation discussed above, whose transcripts should be considered the most reliable or accurate? Rather than privileging one interpretation over another, we advocate for the preservation and archiving of all versions of transcripts. Each version reflects a distinct, yet equally valid, understanding of a given text and may contain insights that would otherwise be lost. By “all versions of transcripts”, we intend the initial draft the researcher produced, any interpretations that linguists may find surprising and perplexing, and even transcripts in which a speaker’s interpretation diverges from the recorded audio.

We further recommend consulting with the original speaker—ideally sooner rather than later.16 This suggestion contradicts Jung and Himmelmann’s (2011: 206) assertion that “[i]t is not necessary, and in fact often not desirable, to work on transcription with the speaker who appears in the recording”. Their reasons include that these speakers may amend their recorded utterances and may not “[restate] precisely what was actually said”. However, these issues hold true for any speaker, and Jung and Himmelmann (2011: 207–218) even provide examples of interpreters (intentionally) deviating from the original recording (see also Chelliah 2001; Dobrin 2021; Crowley 2007; Haviland 1996; Urban 1996).

There are compelling reasons to involve the original speaker in the interpretation of their own recording. Chief among them is the preservation of the speaker’s voice—defined by Webster (2013: 136) as the “ability to express something in a satisfying manner using all of one’s expressive options”. When the researcher transcribes only the editorializations of other speakers, they risk obscuring or erasing aspects of the recorded speaker’s voice. For instance, in the examples featuring Charles’s recorded narratives, Felice rendered ammaaítchiluua ‘store’ as ammaaíschiluua. Had we only recorded Felice’s interpretation, we would have inadvertently marginalized Charles’s voice. Although Felice’s interpretation provides valuable insights into inter-speaker variation, it is only through re-engaging with Charles that we were able to maintain Charles’s agency to express himself.

More broadly, these examples highlight the contribution of audience design to the shape of the narratives we recorded. As Bell (1984) explains, audience design defines how speakers organize their speech style to reflect the audience that they assume. In this case, the audience may include both physically present addressees and imagined future listeners. Speech is never produced in isolation—as Dobrin (2012: 22) writes, “[s]peakers are never just talking; they are always also representing themselves to someone”. Schwartz (2021) similarly observes that speakers being recorded are often aware that they serve as representatives of their speech community, embodying not only their own identity but also their culture.

Although misunderstandings can be resolved by asking speakers for additional clarification, this is not always feasible when playing back a recording for a community collaborator to interpret. The speaker whose voice is featured in the recording may not be able to be physically present—they may be unavailable or may have since passed on. Even if the person is available, if significant time has passed since the creation of the initial recording, the speaker may no longer recall their intended meaning. Additionally, speakers might wish to modify what they originally said to better represent themselves at that particular moment in time. These and other factors compound to make the interpretation process prone to some of the types of variation we observed in our transcripts.

Today, only a few people are literate in Apsáalooke. We do not have the luxury of obtaining written interpretations from a wide range of speakers, nor do we have the resources to train speakers and active listeners (also commonly referred to as passive speakers) to transcribe in an orthography, as is more common in endangered language contexts in Amazonia (Woodbury and England 2004) and Indonesia (Florey and Himmelmann 2009). Even if we can train community members, they may inadvertently alter the content of the recordings to better reflect their vision of how their language and culture should be represented, without much regard for how the original speaker(s) wished to be depicted. This is an occurrence we have commonly observed when working closely with speakers, as they sometimes express a preference to phrase things differently or comment on the ‘incorrectness’ of other speakers’ utterances (see also Dobrin 2021; Chelliah 2001; Haviland 1996; Jung & Himmelmann 2011; Urban 1996; Weber 2021).

Furthermore, working with or training speakers to transcribe and interpret recorded narratives requires resources that could be allocated elsewhere. If a linguist is sufficiently familiar with the language and works independently to transcribe and translate the narratives, that individual might be able to avoid the obstacles arising from conflicting interpretations. However, they may still inadvertently erase and marginalize the original speaker’s voice, since even the linguist’s biases can influence the interpretation process. Even minor decisions, such as how to identify sentence boundaries and whether to include false starts, can alter how the speaker’s voice is conveyed in the transcript. In language revitalization contexts, such transcripts could be viewed as authoritative even though they were created without input from the original speaker and—assuming the linguist is not fluent in the language they are documenting— other speakers of the language.

In the early days of transcribing spoken discourse, before the advent of audio recording technology, transcripts were recorded entirely through dictation, which necessarily precluded the transcription of fast-speech phenomena. Today, written texts from that period are held in high regard, as they are often believed to contain more ‘authentic’ language, especially for languages without known fluent speakers. Some language reclamation practitioners design language classes around such texts (Ko, n.d.). In one example, texts transcribed and written down by James O. Dorsey (1848–1895) in Kanza, a Siouan language, are afforded the same authority and credibility as the speech of a contemporary fluent speaker (for a discussion on the issues surrounding Dorsey’s Kanza materials, see McBride 2009, 2020). When language resources are extremely limited, it behooves one to use whatever materials can be found. Still, caution must be taken when working with past documentation. It is best to approach it with a critical lens.

In each of the case studies presented above, we made implicit and explicit judgments on the various interpretations we received. A reviewer wonders “what transcripts are to be documenting (roughly, production vs. reception)” and “if any criteria are used regarding when to take a stance about which variant to use vs. when to keep multiple variants in parallel”. Ultimately, we believe it depends on one’s research questions and goals. Although we suggest maintaining and archiving all transcripts, we presented only a single speaker’s interpretation on our storytelling mapping platform. We selected each interpretation based on how closely it aligned with the voice of the recorded speaker. For example, we selected the interpretation obtained in person rather than the one created during remote communication over Zoom. However, in this paper, we presented multiple variants in parallel since our goal was to foreground the various interpretations produced by even a single individual.

Of course, not all researchers will want or be able to focus on the quality of the transcripts or their faithfulness to the recordings as we have here. A scholar analyzing naturalistic recordings of conversations or child-caregiver interactions has the additional challenge of transcribing speech amidst background noise, which often results in unintelligible utterances and difficulty in identifying the turns of each speaker. As Amalia Skilton (personal communication, 2024-07-29) notes, “there will often be multiple plausible hypotheses about what is being said or should have been said (for a child)”. Depending on one’s research questions and methods, it is important to consider whether the quantity or quality of the transcripts should be prioritized, as well as to weigh the potential tradeoffs of each approach.

Just like all transcripts, our transcripts are idealizations of spoken discourse. The audio format inevitably excludes all context, including gestures, space, and posture. Video recording captures only slightly more extralinguistic information. Creating texts out of recorded speech inherently results in a degree of decontextualization (Bauman & Briggs 1990). As individuals interpret the recordings, recontextualization also occurs, leading to the (re)creation of ostensibly limitless interpretations. In this sense, the interpretative act will differ from person to person. These points build on other scholars’ calls for documentary linguists to view their work in an interpretivist paradigm, rather than a positivist one where replication and reproducibility are the goals (e.g., Dobrin 2021; Dobrin and Schwartz 2021; Seidel 2016).

Recent efforts have focused on the development of speech recognition (e.g., Prud’hommeaux et al. 2021) and morphological processing technologies (e.g., Eskander et al. 2019) for endangered (or low-resources) languages to overcome the so-called ‘transcription bottleneck’ and expedite the transcription process. If we already obtain a variety of interpretations from different speakers, how can we expect machines, even with input from linguists, to reach a similarly fine-grained level of performance? Even NLP (or natural language processing) tools for high-resource languages, such as English, have difficulty dealing with variation (for an overview of the potentially negative impacts of language technologies, see D’Arcy & Bender 2023). The issue of variation is further exacerbated in endangered language contexts because speakers in a given community may have very different language acquisition histories.

At the start of this paper, we presented two types of transcripts—a naturalized transcript and a denaturalized transcript. Our decision to format the transcripts in these ways was based on our goals: to serve both the community’s language revitalization and reclamation efforts and the work of any scholars interested in researching the language. At the same time, our approach has consequences that may contravene the goals and aspirations of some members of the community. Significant among these is our use of the standard Apsáalooke orthography, which often fails to denote variation. We also made a conscious decision to include only some paralinguistic features in our transcripts, leaving out expressive features such as pausing and prosody. Our transcripts therefore do not fully showcase the speaker’s voice and instead risk reinforcing a Western bias that privileges referential content over expressive or affective dimensions of speech (Webster 2013; see also Kroskrity 2013). However, by presenting more than one transcript of the same stretch of speech, we aim to contribute to what McDonnell (2018) refers to as facilitative documentation: work that is not driven by a specific research question but is designed to support a range of future inquiries, for both the researcher and the community.

The iterative and evolving nature of transcription presents many opportunities for academic and community researchers alike, but only if evidence of this evolution is preserved. What makes language documentation distinct from other disciplines that generate transcripts is the integral role of the archive (Henke & Berez-Kroeker 2016). Based on our experience conducting archival research, it is common to find correspondence and drafts of (un)published works in the collections of earlier scholars. However, there is a very real danger of losing these documents to natural disasters. Today, digital archiving presents other challenges. Unless printed out, internet correspondence, specifically email, is rarely, if ever, deposited, even though born-digital transcripts—much like their handwritten counterparts—may themselves build on previous sets of transcription (e.g., Debenport 2015; Evans and Sasse 2007). In the digital space, transcripts can be overwritten with the click of a button, leaving no trace of these earlier versions. Thus, we encourage researchers to keep track of changes they may make to transcripts, to prevent the loss of that information.

We also recommend that language archivists adopt incremental archiving, a process by which digital files are uploaded in stages, and subsets are accessioned as they become ready. This model is currently implemented by the California Language Archive (CLA; Garrett et al. 2019). In contrast, other archives require that all files in a deposit, along with complete metadata, be submitted in a single submission. While most, if not all, archives allow modifications to existing metadata, not all allow revision to the content of already accessioned material. For example, the CLA does not allow amendments to accessioned content, including adding to, amending, or removing files from a file bundle, thereby ensuring that previous work is not overwritten.17

The first aim of this paper was to discuss and reflect on the challenges of our transcription process. We provided two transcripts, formatted in a naturalized and denaturalized style; the former is more reminiscent of standardized, written discourse, whereas the latter preserves more elements found in spoken discourse, including processes associated with connected speech, hesitation particles, and false starts. The second aim was to show how a single recorded narrative can give rise to a multiplicity of interpretations due both to the means of communication and the subjective nature of the interpretive act itself. We also identified multiple points of variation that scholars of Apsáalooke may wish to explore in the future. We proposed that all drafts of transcripts should be maintained and preserved, including the researcher’s initial drafts and those produced in collaboration with speakers that may not exactly align with the recording.

In fact, many of the narratives about personal experiences were documented without explicit connection to any specific research question. Thus, these texts have the potential to both benefit the speech community and facilitate the research of future scholars (McDonnell 2018). For instance, most are considered the standard genre used in sociolinguistic research on variation as they contain extended stretches of speech recounting personal experiences in both Apsáalooke and English (see Sneller & Barnhardt 2023). Despite the fact that significantly more endangered language data has been recorded, research on sociolinguistic variation remains understudied (Seifart et al. 2018; see also Hildebrandt et al. 2017); these texts could enable future studies into this and other topics.18 That said, we conducted interviews without asking follow-up questions, which may preclude the effects of speaker accommodation and rule out opportunities for backchanneling in the language. As Skilton (2017: 111) remarks, recording this back-and-forth could have provided valuable demographic information about the participant, information that I have not been documenting as thoroughly and would be necessary for any variationist study.19

As a field, we need to more explicitly acknowledge the kinds of variation we encounter and document in our work—variation that, we believe, is not unique to our own Apsáalooke project but likely observed by other language documentarians. By detailing our transcription process, the challenges we faced, and the strategies we employed to address them, we hope to contribute to broader conversations about the transcription of spoken discourse in language documentation. To conclude, we offer a set of considerations that field linguists may find useful when developing transcripts: First, consider your readership and your research goals. As we noted at the outset of this paper, our transcription process was shaped by two primary goals, and we designed our transcripts accordingly. Second, think about providing multiple transcript formats. In our project, we provided two versions to accommodate the needs of different users. Third, be attentive to whose voice—and whose ear—is being privileged in your transcripts. Conflicting interpretations from different individuals are to be expected; when they arise, consider preserving all of them and clearly attributing each interpretation to the individual who provided it. Fourth, keep in mind that standard orthography, while useful for accessibility, can erase socially meaningful variation. Efforts to standardize should be weighed against the risk of introducing negative indexical associations to non-standard forms. Finally, reflect critically on the biases you bring to the transcription process. All researchers carry assumptions and predispositions, and it is essential to recognize and account for them in your work. By taking these considerations into account, we hope that researchers in language documentation will be better equipped to document and represent interpretive variation in future work.