Offline speech systems
VoxRelay
The tested audio foundation for an offline multilingual subtitle system—carefully separating what works now from the larger product still to be built.
VoxRelay is intended to turn local speech into English subtitles without depending on a network connection. The completed milestone is the audio front end; streaming recognition and translation remain future work.
Can the low-level audio pipeline be made numerically trustworthy before adding streaming ASR, translation and a subtitle interface?
My contribution
- Built WAV loading, resampling, log-Mel extraction, a small temporal encoder, packaging and tests.
- Specified the 16 kHz mono, 512-point FFT, Hann STFT and 80-filter HTK Mel pipeline.
- Compared features numerically against torchaudio and merged the foundation milestone to main.
Signature view
Trace the evidence.
Move across the research chain. The case study below remains complete and readable without this interaction.
Question
Trust the audio representation first
A subtitle system cannot be evaluated cleanly if its resampling and feature extraction are inconsistent.
Method, as a sequence of decisions
- 01
Load and normalise waveform audio to 16 kHz mono.
- 02
Compute a Hann-windowed STFT with a 512-point FFT.
- 03
Project magnitudes through 80 HTK Mel filters and take natural-log features.
- 04
Pass [frames, 80] features into a small temporal encoder for the next research stage.
What the record actually supports
Audio target rate
Mono input normalised for the downstream speech pipeline.
Feature channels
HTK Mel filters with natural-log output shaped [frames, 80].
Maximum torchaudio difference
Recorded numerical comparison for the completed feature extractor.
Evidence: September 2026 development record and October CV. The completed scope is the audio foundation only.
Choices and trade-offs
Why start below the model layer?
Small inconsistencies in resampling, windowing or Mel conventions can make later ASR comparisons misleading.
Why remain offline?
Local processing can support privacy, resilience and low-connectivity settings, but it raises tighter compute and latency constraints.
What has not been established
- Streaming ASR, translation, subtitle rendering and Android delivery are not yet implemented.
- The numerical feature comparison is not a word-error-rate or end-to-end latency result.
- No public repository URL was recovered for the current implementation.
What I would investigate next
- Q1Which compact ASR model best balances multilingual recall, memory and streaming delay?
- Q2How should partial hypotheses be stabilised before translation?
- Q3What latency budget is acceptable for readable live subtitles on commodity hardware?
No public paper or repository is attached yet. The case study is intentionally limited to the evidence in the supplied record.
Peer notes / VoxRelay
See a missed assumption?
Research suggestions, methodological critique and relevant collaboration ideas are particularly useful.