mustafaoksuz/soztr

SÖZTR: Turkish Conversational Speech Transcript Corpus

Status: Early development / prototype
Language: Turkish (tr)
Data modality: Text transcripts (no audio is included in the inspected examples)
License status: Unverified for the inspected source recordings and transcripts; not cleared for public redistribution

Overview

SÖZTR is a developing research corpus of Turkish-language speech transcripts from interviews, television discussions, news reports, and related spoken-media formats. Its goal is to preserve examples of naturally occurring spoken Turkish in machine-readable form for research in natural language processing (NLP) and computational linguistics.

According to the project workflow, source recordings are converted to text using Whisper-family automatic speech recognition (ASR) models. The transcripts may then be represented as timestamped plain text or structured JSON. Only one of the three inspected samples explicitly identifies its ASR model; the remaining files do not yet contain model provenance.

The corpus is not currently a gold-standard, manually validated transcription benchmark. Its transcripts and speaker labels should be treated as automatically generated and potentially erroneous unless a record explicitly documents human verification.

Türkçe özet

SÖZTR, röportajlar, televizyon tartışmaları, haber röportajları ve benzeri yayınlardan elde edilen Türkçe konuşma transkriptlerini derlemeyi amaçlayan bir araştırma projesidir. Ses kayıtlarının metne dönüştürülmesinde Whisper ailesinden otomatik konuşma tanıma (ASR) modelleri kullanılmaktadır.

Amaç, doğal konuşma Türkçesine ait soru-cevap yapıları, dilsel tekrarlar, duraksamalar, gündelik ifadeler ve farklı dönemlerdeki konuşma biçimlerini araştırılabilir bir veri yapısında korumaktır. Transkriptler ve konuşmacı etiketleri, ayrıca doğrulanmadıkça hatasız veya elle kontrol edilmiş kabul edilmemelidir.

Intended Research Uses

Depending on source permissions and the quality of individual records, SÖZTR may support:

• Research on Turkish spoken and conversational language.
• Dialogue structure and turn-taking analysis.
• Domain-specific language modeling and exploratory NLP research.
• Study of colloquial expressions, hesitations, repetitions, and discourse markers.
• Historical and contemporary spoken-Turkish comparisons, where recording dates are verified.
• Research on transcription errors and speaker segmentation, when independently checked reference material is available.

Not suitable without additional validation: ASR accuracy benchmarking (WER/CER), speaker-diarization benchmarking, fact-checking historical statements, or evaluation requiring verified ground truth. No verified train/validation/test splits are provided in the current sample.

Sources and Coverage

The intended source categories include:

• Interviews (interview): one-to-one or moderated interviews.
• Debates (tv_debate): multi-speaker television or panel discussions.
• News reports (news_report): reports combining narration, questions, and short participant responses.
• Other spoken media: podcasts, public discussions, and similar material, if duly licensed.

These are target categories, not a claim that every category is already represented in a released dataset.

Inspected sample inventory (13 October 2026)

The following figures describe three locally inspected example files, not a public release or a representative sample of Turkish speech. Word counts use whitespace splitting and are approximate.

File	Format	Current category	Observed units	Speaker labels	Approx. words	Provenance available in file
TR000001.txt	Timestamped TXT	interview	10 timestamped sections	2	580	YouTube title, URL, timestamp field, ASR model
TR000002.json	JSON	interview	46 dialogue entries	2	3,105	No recording URL/date/model field
TR000003.json	JSON	interview (reclassification suggested: news_report)	22 dialogue entries	7	627	No recording URL/date/model field
Total			78	Not additive	4,312

The speaker-label counts above do not establish the true number of distinct speakers. Some entries appear to combine speech from multiple participants. TR000003 contains references to 1973–1974, but these references do not independently establish the broadcast date.

Current Data Format

The two inspected JSON files follow this structure:

{
  "id": "EXAMPLE0001",
  "language": "tr",
  "category": "interview",
  "dialogue": [
    {"speaker": "A", "text": "Merhaba, bugün hangi konuyu konuşacağız?"},
    {"speaker": "B", "text": "Bugün şehir yaşamı hakkında konuşacağız."}
  ]
}

Illustrative example written for this card; it is not a quotation from a source recording.

The observed JSON top-level fields are:

Field	Type	Meaning
id	string	Locally assigned record identifier (TR000002, etc.).
language	string	Transcription language (tr).
category	string	Broad content category; may need review.
dialogue	list of objects	Sequence of {speaker, text} entries.

TR000001.txt instead contains a header with source information and timestamped transcript lines. The inspected TXT and JSON formats are not yet harmonized.

Recommended Unified Schema (Planned)

A future normalized release should use JSONL, with one recording/conversation per line, and should preserve raw transcriptions separately from corrected versions. The following is a proposed schema, not an existing field specification:

{
  "id": "EXAMPLE0001",
  "language": "tr",
  "category": "interview",
  "source": {
    "type": "youtube",
    "title": "Example interview",
    "publisher": null,
    "url": null,
    "broadcast_date": null,
    "rights_status": "pending_review"
  },
  "transcription": {
    "method": "asr",
    "model_id": null,
    "processed_at": null,
    "verification_status": "unverified"
  },
  "segments": [
    {
      "start": null,
      "end": null,
      "speaker_id": "A",
      "text": "Merhaba, bugün hangi konuyu konuşacağız?"
    }
  ]
}

The example is synthetic. Null values explicitly mean that the information is unavailable or unverified.

Recommended rules:

1. Record the actual recording or broadcast date, transcription date, and collection date in different fields.
2. Document the exact ASR model ID and version, language setting, and important inference/processing parameters when known.
3. Keep a source URL, source publisher, and record-specific rights status. A URL alone is not evidence of a license.
4. Preserve raw ASR text and, if produced, a separately versioned human-corrected transcript.
5. Distinguish actual speaker identity from an arbitrary A/B label; unknown or unverified speakers should remain explicitly unknown.
6. Retain segment time boundaries when available; do not invent timestamps for untimed transcriptions.
7. For segment-level modeling, offer a separate utterance-level export with recording ID, speaker label, and text.
8. Track corrections, schema changes, dataset releases, and any removed records.

Collection and Transcription Methodology

The intended processing pipeline is:

1. Identify recordings and document their origin and rights status.
2. Obtain and process recordings only when the intended use is legally permitted.
3. Produce Turkish ASR transcripts using a documented Whisper-family model.
4. Retain original ASR output and, where available, segment timestamps.
5. Detect or annotate speaker transitions separately; ASR output alone does not validate speaker identity.
6. Normalize to a consistent machine-readable structure while preserving an original-text version.
7. Run quality checks (JSON parsing, required fields, duplicates, unusually long turns, missing provenance, transcription anomalies).
8. Conduct sampled human review against the original recording where access and permissions allow.
9. Release only materials with verified authorization and appropriate privacy safeguards.

Known Issues in the Inspected Samples

• Inconsistent formats: one TXT and two JSON files.
• Missing provenance: recording source metadata is absent from the inspected JSON files.
• Missing timing: JSON dialogue entries do not contain timestamps.
• Possible speaker merging: some dialogue entries seem to contain more than one person's speech.
• Oversized segments: one TXT section contains approximately 479 words; one JSON turn contains approximately 345 words.
• ASR artifacts: suspected word substitutions, repeated fragments, and irregular punctuation appear in the texts.
• Potential misclassification: the archival traffic report is currently tagged interview; news_report may be more accurate.
• Unverified history: mentions of calendar years inside a transcript do not prove recording or broadcast dates.
• Missing quality metrics: no verified WER/CER, diarization accuracy, or inter-annotator agreement is available.

Quality, Bias, and Limitations

The current examples appear to be drawn from journalistic or broadcast contexts. They do not establish representativeness across Turkish regions, generations, socioeconomic groups, dialects, or everyday private conversation. Historical and contemporary source material may differ substantially in topic and language use.

Recognition errors may disproportionately affect proper names, dates, older recordings, accents, rapid speech, overlapping speech, and domain-specific terminology. A speaker label represents only the available annotation, not a verified identity. Statements within transcripts are the views or words of recorded participants and must not be interpreted as dataset maintainers' endorsements or verified facts.

This text-only collection does not include the original audio in the inspected examples. Without recordings and checked reference transcripts, claims about ASR transcription accuracy cannot be made.

Privacy, Copyright, and Release Policy

Current metadata license: unknown. The inspected samples do not provide evidence that rights holders have authorized open redistribution of their full transcripts. They should not be published publicly as a CC BY 4.0 dataset until the relevant rights and privacy questions are resolved.

• Public availability of a television or YouTube recording does not automatically permit redistribution of the entire transcript.
• Transcribing audio with Whisper does not remove underlying intellectual-property rights.
• Attribution alone is not a substitute for permission where permission is required.
• Personal information and sensitive information must be reviewed before any release; public-figure status does not remove all privacy considerations.
• Record-by-record licensing and access restrictions should be documented.
• Content for which redistribution rights cannot be verified should remain outside any public data release; making a repository private does not by itself establish processing rights.
• If and when eligible records are authorized for CC BY 4.0 distribution, publish those records in a clearly identified, licensed subset. Do not apply that license to uncleared third-party material.

Relevant guidance: Turkish Ministry of Culture and Tourism — Copyright Directorate.

Repository Structure (Suggested)

soztr/
├── README.md                 # This dataset card
├── docs/
│   ├── annotation_guidelines.md
│   ├── sources_and_rights.md
│   └── changelog.md
├── examples/
│   └── schema_example.json   # Original, synthetic illustrative record
└── data/
    └── cleared/              # Only recordings/transcripts authorized for redistribution
        └── conversations.jsonl

Do not upload uncleared recordings or full transcripts to a public repository. A separate private working environment may be used for appropriately authorized processing and review.

Once data/cleared/conversations.jsonl actually exists, the README YAML metadata can be extended with a Hugging Face configs entry pointing to that file. No split or subset configuration is asserted for files that have not yet been released.

Versioning and Citation

The project is at the prototype stage. No externally validated corpus release, persistent DOI, reference paper, or stable train/test benchmark is claimed in this card.

When a publishable version becomes available, the maintainers plan to provide release-specific statistics, source and rights inventories, a changelog, and a recommended citation with a version identifier and (if available) a DOI.

Contributions and Corrections

Contributions may include independently authorized transcripts, provenance metadata, transcription corrections, annotation guidelines, and validation tools. Contributions must not include materials that the contributor lacks rights to share. Suspected transcription errors, attribution issues, and rights concerns should be reported through the dataset repository's maintainer contact or issue process without reposting sensitive material.

---

Documentation note: This card reflects inspection of three example files as of 13 October 2026. It does not certify copyright clearance, recording authenticity, speaker identity, transcription quality, or availability of a public data release.