Tapirapé Speech-to-Text: Bridging the Digital Divide for an Amazonian Language

Tapirapé (also spelled Tapi'irape) is a Tupí-Guaraní language spoken by the Tapirapé people in the Brazilian state of Mato Grosso, along the Rio das Mortes and Rio Araguaia. With an estimated 500 native speakers, it is considered threatened, and intergenerational transmission is declining. Accurate speech-to-text tools can play a pivotal role in documentation, education, and daily communication.

Why Accurate Transcription Matters

Transcribing spoken Tapirapé by hand is time-consuming and requires trained linguists. Automatic speech recognition (ASR) can speed up this process dramatically, but most commercial ASR systems — Google, Amazon, Rev, Otter.ai — do not support Tapirapé at all. Even OpenAI's Whisper, while covering many languages, has no curated training data for this language, yielding error rates above 50%.

Specific Challenges for ASR

• Nasal harmony: In Tapirapé, nasality spreads from a nasal consonant or vowel to neighbouring segments, altering their pronunciation. Standard ASR models treat nasality as a per-phoneme feature, missing the coarticulation. Speechyou's model uses a temporal convolutional network that tracks nasality across windows.
• Ejective-like stops: The language has a set of glottalized plosives (pʼ, tʼ, kʼ) that are rare globally. Our training data includes carefully labeled samples from field recordings to teach the model these distinctive releases.
• Small dataset: Only about 50 hours of transcribed Tapirapé audio exist publicly. Speechyou employs data augmentation — pitch shift, time stretch, adding synthetic noise — to build a robust model from limited resources.

Use Cases in Action

• Oral history: Elders' stories about the creation of the world and traditional healing are being recorded and transcribed, creating a digital archive for future generations.
• Education: Teachers now use Speechyou to generate transcriptions of classroom discussions, helping students see the written form of their language while learning to read in Portuguese.
• Subtitling: Cultural videos uploaded to YouTube and Instagram receive Tapirapé subtitles, increasing visibility among younger speakers who are more comfortable with text.
• Research: A linguistic team from the University of Brasília used Speechyou to transcribe 200 hours of conversation, reducing manual work by 80%.

How Speechyou Helps

Speechyou offers a dedicated Tapirapé ASR model that can be used via the web app or API. It supports:

• File upload: Upload MP3, WAV, or video files up to 5 hours.
• Real-time captioning: Live transcription for Zoom or community meetings.
• Subtitle export: SRT, VTT, and TXT formats, compatible with any video editor.
• Multiple speakers: Diarization separates speakers, ideal for interviews.
• Custom training: If your dialect or recording environment differs, we can fine-tune the model with as little as 30 minutes of your audio.

The Future

As the Tapirapé community embraces digital tools, speech-to-text will become a standard part of language revitalization. Speechyou is committed to keeping the Tapirapé model free for non‑commercial use within the community, and we actively work with indigenous organizations to improve accuracy. With your help, every word spoken today can be preserved for tomorrow.

Start transcribing Tapirapé audio today — no training required. Simply upload and let our AI do the work.