WARDEN: Endangered Indigenous Language Transcription and Translation with 6 Hours of Training Data

Technical Summary: WARDEN – Endangered Indigenous Language Transcription and Translation with 6 Hours of Training Data

Problem Statement
The documentation of endangered languages, such as Wardaman (an Australian Indigenous language with only two full speakers as of 2025), is a labor-intensive process requiring linguists to manually produce time-aligned transcriptions and translations. While automatic speech recognition (ASR) and machine translation (MT) offer potential acceleration, standard deep learning approaches are infeasible for such low-resource scenarios. Existing models typically require tens to hundreds of hours of annotated data to achieve acceptable performance, whereas the available Wardaman dataset consists of only 6 hours of time-aligned audio-transcription-translation pairs. Furthermore, unified models that attempt to perform both transcription and translation simultaneously often fail in extremely low-data regimes due to the lack of sufficient training signal.

Methodology
The authors propose WARDEN (Wardaman Decoding ENgine), a two-stage framework designed to decouple transcription and translation, thereby enabling effective learning with minimal data.

1. Transcription Stage (ASR):

   • Model: The system fine-tunes the Whisper-large-v3 model.
   • Cross-Linguistic Initialization: To address the scarcity of Wardaman data, the authors identify a phonetically similar "proxy" language to initialize the model's language tokens. Using the PHOIBLE database, they computed Hamming distances between Wardaman's phoneme inventory and six candidate languages (Sundanese, Uzbek, Hausa, Croatian, Occitan, English).
   • Selection: Sundanese was identified as having the smallest phoneme distance. The model is initialized using the Sundanese language token (<su>) before fine-tuning on the 6 hours of Wardaman data. This strategy aims to transfer phonological knowledge and accelerate convergence.

2. Translation Stage (MT):

   • Challenge: Directly fine-tuning large language models (LLMs) or ASR-based translators on 6 hours of data yields poor results due to overfitting and lack of domain-specific knowledge.
   • Lexicon-Augmented Workflow: The authors compile a Wardaman-English dictionary (approx. 2,300 entries) from expert annotations.
   • Lexicon Matcher: Before translation, a rule-based matcher retrieves relevant dictionary entries for words in the ASR transcript. Matching is performed via:
     • Character Error Rate (CER): Matching transcription words to dictionary entries with a CER threshold ($\tau = 0.2$).
     • Affix Matching: Positional matching of prefixes and suffixes to handle derived forms.
   • LLM Integration: The retrieved lexicon entries (formatted as word, CER, part-of-speech, and gloss) are concatenated with the ASR transcript and fed into an LLM (specifically Qwen3-8B).
   • Training: The LLM is fine-tuned using Low-Rank Adaptation (LoRA). The input includes the transcript and the matched lexicon cues, transforming the LLM into a knowledge-grounded interpreter rather than a purely data-driven translator.
   • Data Augmentation: The training data is augmented by concatenating short utterances into longer sequences and by using Whisper's own noisy transcription outputs as input to simulate real-world ASR errors.

Key Contributions

• Two-Stage Architecture: The paper demonstrates that separating transcription and translation is superior to unified approaches in extremely low-resource settings (6 hours of data).
• Phonetic Proxy Initialization: The introduction of initializing ASR models with tokens from a phonetically similar language (Sundanese) to overcome data scarcity.
• Lexicon-Grounded Translation: A novel pipeline that integrates expert-compiled dictionaries directly into the LLM inference and fine-tuning process, allowing the model to leverage external knowledge rather than relying solely on internal parameters.
• Baseline Establishment: The creation of a strong baseline for Wardaman processing using only 6 hours of annotated data, outperforming larger open-source and proprietary models.

Experimental Results

• Transcription: The WARDEN system (Whisper fine-tuned with Sundanese initialization) achieved a Word Error Rate (WER) of 0.52. This outperformed standard Whisper fine-tuning (0.64) and zero-shot baselines (1.62). The use of the Sundanese token reduced WER by 0.12 compared to standard fine-tuning.
• Translation: Using the Qwen3-8B model with lexicon conditioning and fine-tuning, WARDEN achieved a BLEU-4 score of 12.40.
  • This significantly outperformed the best GPT-5 result (7.54) and other fine-tuned baselines.
  • Ablation studies showed that removing either the lexicon or the fine-tuning reduced the BLEU score by 6.28 and 9.57, respectively.
  • The system also demonstrated robustness, maintaining performance even when fed noisy ASR outputs.

Significance and Claims
The paper claims that WARDEN successfully escapes the "curse of limited data" that hinders existing systems. By leveraging linguistic insights (phonetic similarity) and expert knowledge (dictionaries), the system achieves high performance with a fraction of the data typically required. The authors position WARDEN not as a replacement for linguists, but as a tool to accelerate the documentation process, reducing the time commitment from days to hours for processing audio recordings. The work emphasizes that such tools can support community-led language revitalization, provided they are developed with respect for the data ownership and preferences of the speech community. The authors explicitly state that the effectiveness of the design is verified by its ability to outperform larger, data-hungry models in this specific low-resource context.