Qualitative researchers often need reliable, low-cost transcription on mobile devices to capture interviews and field notes. The primary keyword for this guide is "termux stt" and this post explains how the new termux-stt package (published on PyPI) makes on-device speech-to-text and speaker diarization feasible on Android Termux. The payoff: actionable benchmark numbers, installation patterns, and concrete tips to feed high-quality transcripts into AI-enabled qualitative analysis pipelines.
Key Takeaways
termux-stt is an on-device speech-to-text and diarization framework published on PyPI, and the PyPI page includes empirical benchmarks and package metadata for the Aug 20, 2026 release.
- The PyPI package upload date is Aug 20, 2026 and the release includes source and wheel files sized 47.9 kB and 51.5 kB respectively, according to the PyPI file metadata.
- According to the PyPI page on Aug 20, 2026, a live-audio benchmark processed a 37.91 second input in 32.79 seconds (RTF 0.865x) using whisper.cpp Base on a Samsung Galaxy A35.
- The PyPI empirical benchmarks on Aug 20, 2026 report memory footprints ranging from ~100 MB for Vosk small to ~1.5 GB for whisper.cpp medium, and the hybrid Vosk+Whisper pipeline runs at about ~350 MB with reported accuracy around 92% on Galaxy A35 tests.
- The termux-stt project page states, "termux-stt is an all-in-one, production-ready speech-to-text and speaker diarization framework engineered natively for Android Termux (ARM64 / aarch64)."
What termux-stt does and how the on-device STT works
termux-stt provides a unified API to run whisper.cpp, Vosk, Sherpa-ONNX, or a hybrid Vosk+Whisper pipeline on Android Termux devices, according to the PyPI project page.
According to the PyPI page, termux-stt uses subprocess isolation to run C++ inference binaries so that native crashes do not kill the host Python process, and it normalizes input audio to single-channel 16 kHz PCM with ffmpeg as a preprocessing step.
According to the PyPI page, the library supports file transcription, real-time microphone streaming, and hybrid speaker diarization that combines Vosk 128d X-Vector fingerprints with Whisper segments to yield aligned speaker labels without PyTorch.
According to the PyPI page, the Python SDK is a simple three-step flow: create_engine(...), result = engine.transcribe(file), and then export using result.to_srt() or result.to_json().
Findings Snapshot
| Date | Metric | Value | Implication |
|---|---|---|---|
| Aug 20, 2026 | Package upload date | Aug 20, 2026 | Release metadata and files published on PyPI |
| Aug 20, 2026 | Live-audio processing | 37.91s input → 32.79s processing (RTF 0.865x) | On-device whisper.cpp Base runs faster than real time on Galaxy A35 in this test |
| Aug 20, 2026 | Model memory footprints | Vosk small ~100 MB; whisper.cpp tiny ~150 MB; whisper.cpp medium ~1.5 GB; hybrid ~350 MB | Choose model by RAM budget on Android devices |
| Aug 20, 2026 | Hybrid accuracy | Hybrid Vosk+Whisper reported ≈92% accuracy in Galaxy A35 benchmarks | Hybrid pipeline balances diarization and transcript quality for interviews |
| Aug 20, 2026 | Files | termux_stt-1.0.0.tar.gz (47.9 kB), termux_stt-1.0.0-py3-none-any.whl (51.5 kB) | Small Python wrapper with native binary provisioning |
Implications for qualitative researchers
On-device transcription with termux-stt reduces data sovereignty risk by keeping raw audio and initial transcripts on the device, according to the PyPI project description and its design goals.
According to the PyPI benchmarks on Aug 20, 2026, teams with Android devices similar to a Samsung Galaxy A35 (6 GB RAM) can run a hybrid diarization pipeline in roughly 1.45 RTF and expect usable transcripts with speaker labels for interviews, which changes how fieldwork logistics and consent should be handled.
According to the PyPI troubleshooting notes, termux-stt automates common Android packaging issues such as extracting libvosk and handling ffmpeg normalization, which reduces setup time compared with manual cross-compilation workflows.
According to the PyPI page, hybrid diarization avoids PyTorch OOM crashes on mobile by keeping heavy models off-device and using pure Python clustering where possible, making longer interviews more feasible on constrained hardware.
How Evidano Helps
What Evidano is and why it matters here
Evidano is an AI-powered qualitative data analysis platform that helps researchers analyze interviews, open-ended surveys, and documents.
Evidano ingests transcripts, timestamps, and speaker labels exported from mobile STT tools such as termux-stt and converts them into thematic, frequency, and cross-segment analyses that speed synthesis for research teams.
For secure projects, Evidano encrypts uploaded data and does not use customer data to train third-party models, which complements termux-stt's on-device privacy posture; see the Evidano data security page for details.
Problem: Low-quality or unstructured mobile transcripts → Solution: Automated cleaning
Problem: Mobile transcripts often have inconsistent punctuation, speaker labels, and timing metadata which slows coding and analysis.
Solution: Evidano automatically ingests SRT, RTTM, and JSON outputs that termux-stt exports and normalizes timestamps, merges segments, and preserves speaker diarization for downstream thematic coding.
Problem: Slow synthesis across many interviews → Solution: Thematic + cross-segment analysis
Problem: Teams spend days manually reading hours of interviews collected on Android devices.
Solution: Evidano applies AI-enabled thematic extraction, co-occurrence networks, and cross-segment frequency counts so researchers can prioritize reading where themes cluster; see Evidano features for visualization examples.
Problem: Non-English or mixed-language audio → Solution: Translation + custom dictionaries
Problem: Field interviews often include code-switching and domain terms that reduce STT accuracy.
Solution: Evidano supports translated transcripts and custom dictionaries to reconcile termux-stt language detection and ensure consistent codes across languages; see Evidano translation for details.
FAQ: termux stt
Can I run high-quality diarized transcripts on a midrange Android phone?
Yes, you can get usable diarized transcripts on a midrange Android phone when you choose appropriate models and settings.
According to the PyPI benchmarks published on Aug 20, 2026, a hybrid Vosk+Whisper pipeline ran on a Samsung Galaxy A35 (6 GB RAM) with a reported memory footprint of ~350 MB and produced approximately 92% accuracy in the author’s tests, which indicates practical viability for interview capture.
How do I export speaker labels and subtitles from termux-stt for analysis?
Export speaker labels and subtitles using the included result.to_rttm() or result.to_srt() methods or the CLI exports.
According to the PyPI API summary, the Python SDK returns structured objects with segments you can export to JSON, SRT, VTT, or RTTM and then ingest into qualitative analysis tools like Evidano for coding and cross-segment queries.
What are the recommended termux-stt models for low-RAM devices?
Use Vosk small or whisper.cpp ggml-tiny for the lowest RAM footprint.
According to the PyPI empirical table on Aug 20, 2026, Vosk small used approximately 100 MB RAM and whisper.cpp ggml-tiny used about 150 MB in the Galaxy A35 tests; choose Vosk when very low RAM and fast RTF (0.25) are required.
Is on-device transcription private enough for sensitive interviews?
On-device transcription reduces exposure by keeping raw audio on the device, but research projects must still obtain informed consent and follow institutional policies.
According to the termux-stt project notes on PyPI, the framework is designed for on-device processing; teams should still follow ethics and data governance rules and consider local encryption before moving transcripts off-device.
Conclusion & Next Steps
termux-stt, published on PyPI with files uploaded on Aug 20, 2026, makes on-device STT and diarization practical on Android Termux by unifying whisper.cpp, Vosk and Sherpa-ONNX under a single API.
According to the PyPI benchmarks, choosing model size trades RAM for accuracy and the hybrid pipeline is a pragmatic middle ground for interview workflows on devices similar to a Galaxy A35.
If you collect interviews with termux-stt, export SRT/RTTM/JSON and import them into Evidano to run thematic, frequency, and cross-segment analyses that save synthesis time; start by reviewing Evidano features and the Evidano data security page.
To try it, Try Evidano for free
Topics
- termux stt
- on-device speech-to-text
- termux speech-to-text
- mobile diarization
- whisper.cpp termux
Keep reading
- Commentary on NewsFaster Photovoice Qualitative Analysis with AIHow AI accelerates photovoice qualitative analysis: extract themes, frequencies, and policy-ready insights from community photos and narratives. Learn methods and try Evidano.
- Commentary on NewsAI-assisted qualitative analysis for museum observationsHow AI-enabled qualitative analysis accelerates coding and insight from museum visitor observations. Learn methods and stats from PLOS ONE, and try Evidano.
- Commentary on NewsAI qualitative insights: epilepsy medication adherenceAI-enabled qualitative analysis of epilepsy medication adherence in Uganda: 277 patients, 66.5% adherence, and actionable insights for researchers and care teams.