Voice, Speech & Edge AI — Public Research Library
Updated: 24 October 2026 · Editor: Nauman · Scope: Public technology research, independent of private product IP.
I follow voice technology because speech is becoming a practical interface for phones, computers, wearables and inexpensive microcontrollers. This is an evidence-led reading and experiment library, not a claim that I have tested every model. Vendor claims are labelled as such.
Start here — research and articles
| Topic | What matters | Source |
|---|---|---|
| Cactus Whistle — tiny offline STT | 16.9 MB, seven languages, CPU-oriented, word timestamps and keyword biasing. Urdu is not among the seven advertised languages. | Cactus announcement · model card |
| Whistle porting and fine-tuning | Source for reproducibility, vocabulary, portability and model adaptation. | Cactus engineering articles |
| Whistle on ESP32-S3 | Independent embedded implementation reported; hardware performance and memory claims require reproduction on an actual board. | Implementation reference |
| Phonon — compact on-device TTS | 31M/40M parameter TTS family; investigate memory and audio quality on mobile/SBC hardware rather than assuming ESP32 compatibility. | Release reporting |
| Qwen / CosyVoice TTS (China) | Multilingual voice generation, streaming, voice cloning and dialect capabilities vary by checkpoint and hosting region. | CosyVoice repository · Qwen speech documentation |
| Fun-ASR / Qwen Audio (China) | Chinese dialects, multilingual ASR, local serving and realtime APIs; inspect specific model licenses. | Fun-ASR · QwenAudio projects |
| Granite Speech 5.0 | Open multilingual STT and keyword biasing, local deployment paths; independently test benchmark claims. | Release reporting |
| ESPHome Voice Assistant | Practical local home voice endpoints, wake words, microphones, speakers and Home Assistant integration. | ESPHome voice assistant |
| ESP-SR / Espressif | Official embedded wake-word, command recognition, audio front-end and DSP software. | ESP-SR repository |
| OpenAI audio | Streaming transcription and realtime speech APIs as cloud comparators. | Audio documentation |
Hardware watchlist
ESP32-S3 — microphone capture, VAD, wake words, offline command recognition, and now experimental open-vocabulary ASR. Test boards with PSRAM, I2S microphone (e.g. INMP441), speaker DAC/amplifier, battery and flash requirements. ESP32-P4 — evaluate greater processing headroom and peripheral tradeoffs; do not assume wireless connectivity is built in. Small Linux boards — Raspberry Pi and similar SBCs are useful comparison points for heavier local STT/TTS. Also track Seeed XIAO ESP32S3 Sense, ESP32-S3-BOX and audio-focused development boards, checking revisions and availability before purchase.
What I want to test
- Whistle on ESP32-S3 with real microphones, measuring latency, memory, temperature, power and noisy-room accuracy.
- A completely offline voice loop: wake word → speech recognition → local action → spoken response. Compare TTS that actually fits the device.
- English and Urdu/English mixed speech: model language coverage is a hard gate, not a feature to assume.
- Qwen/CosyVoice and other Chinese open-source models on local hardware; inspect licence, runtime, quality and privacy.
- A speech-to-tool-call demo with explicit confirmation before any consequential action.
Editorial rules
Public articles are original commentary, source-linked, dated, and independent of internal architecture, financials, user data, private benchmarks or unpublished roadmaps. An article should explain the claim, evidence, limitations, real-world hardware implications and a reproducible experiment. No automatic publication of speculative claims.
Related internal context
A separate private Brain Planner radar tracks project relevance. This public library deliberately does not reproduce its private decisions.