Besemah Speech to Text: Preserving a Malayic Language with AI

Besemah (Basemah) is a Malayic language spoken by approximately 400,000 people in the highlands of South Sumatra, Indonesia. The language is part of the larger Malayic subgroup within the Austronesian family, sharing roots with Palembang Malay and Minangkabau. Despite its rich oral tradition, Besemah has limited written resources and is considered vulnerable by UNESCO. Accurate speech-to-text technology can play a vital role in documenting and revitalizing this language.

Why Besemah Speech Recognition Matters

For Besemah speakers, many of whom are bilingual in Indonesian, transcription tools open up new possibilities:

• Preserving oral histories: Traditional stories, epic poems, and ceremonial speeches can be transcribed and archived.
• Creating educational materials: Teachers can generate written texts from spoken lessons for literacy programs.
• Making media accessible: Local radio and YouTube content can be subtitled in Besemah, reaching a wider audience.
• Supporting research: Linguists and anthropologists can transcribe interviews and field recordings with high accuracy.

Challenges in Transcribing Besemah Audio

Besemah poses several unique challenges for automatic speech recognition:

• Vowel length and quality: The language distinguishes between short and long vowels, as well as vowel qualities like /a/ vs /ɑ/. Generic ASR models often miss these distinctions, leading to errors.
• Final glottal stop: Many words end in a glottal stop /ʔ/ that is not written in standard orthography. This can cause ambiguity between words that differ only by the presence or absence of the glottal stop.
• Code-switching: Speakers frequently mix Besemah with Indonesian or Palembang Malay. A robust ASR system must handle multiple languages in a single utterance.
• Limited training data: Besemah has very few transcribed datasets, making it a low-resource language for machine learning.

How Speechyou Handles Besemah

Speechyou is built to handle low-resource languages like Besemah. Our model uses transfer learning from related Malayic languages to achieve high accuracy even with limited initial data. Key features include:

• Dialect adaptation: Users can select their dialect (e.g., Lintang, Kikim) or provide sample audio to fine-tune the model.
• Mixed-language support: The system recognizes code-switching and transcribes each language correctly.
• Customizable output: You can adjust spelling conventions and include diacritics for vowel length.
• Subtitle generation: Output SRT and VTT files directly from audio or video files.

Use Cases for Besemah Transcription

• Community radio: Transcribe broadcasts for written archives and show notes.
• Oral history projects: Record and transcribe interviews with elders to preserve traditional knowledge.
• YouTube subtitles: Add Besemah subtitles to videos about local culture, cooking, or ceremonies.
• Language learning: Create transcripts of spoken Besemah for learners and teachers.
• Research: Process field recordings for linguistic analysis.

Getting Started

To transcribe Besemah audio with Speechyou, simply upload your file and select the Besemah (Latin) language model. The system will process your audio and return a transcript with timestamps. You can then export subtitles in SRT or VTT format, or edit the transcript directly in the web interface. For best results, use clear audio with minimal background noise and specify the dialect if known.

By providing accurate, AI-powered speech-to-text for Besemah, Speechyou helps preserve a valuable linguistic heritage and makes Besemah content more accessible to the world.