faster-whisper: Faster Whisper transcription with CTranslate2
Summary¶
faster-whisper is a CTranslate2-based reimplementation of OpenAI Whisper, achieving up to 4× speedup over the original with reduced memory. [EVIDENCE src-001] It supports int8 CPU quantization, word-level timestamps via word_timestamps=True, and built-in Silero VAD via vad_filter=True. [EVIDENCE src-001] It is YazSes's current transcription backend.
Key Findings¶
- Latency: On CPU (int8), tiny.en transcribes a 3-second chunk in ~300–600 ms on a modern x86 core. [EVIDENCE src-001] base.en is ~2× slower. [HYPOTHESIS]
- Streaming: faster-whisper is primarily a batch engine; real-time streaming requires wrapper projects (whisper_streaming, WhisperLive) that call it on rolling audio windows. [EVIDENCE src-001]
- Word timestamps: The
word_timestamps=Trueparameter returns per-word start/end times. These can be used to position a cursor back in injected text for correction. [EVIDENCE src-001] - VAD: Built-in Silero VAD (
vad_filter=True) suppresses silence, reducing unnecessary decode calls. [EVIDENCE src-001] - Language detection: Automatic language detection on the first chunk. Explicit
language="en"bypasses detection and saves ~40 ms. [EVIDENCE src-001] - No GPU required: int8 CPU inference is the primary design target, aligning with YazSes's no-cloud, no-GPU constraint. [EVIDENCE src-001]
Relevance to YazSes Innovation Domains¶
| Domain | Relevance | Notes |
|---|---|---|
| Streaming ASR | HIGH | Core backend; needs wrapper for partial hypothesis display |
| SSH/remote | MEDIUM | Text output can be forwarded; audio capture stays local |
| Code awareness | LOW | No code-specific training; requires grammar layer on top |
| Accessibility | MEDIUM | Needs atypical-speech fine-tuning; architecture supports it |
| AR/VR/XR | MEDIUM | Must emit text over socket for XR consumers |
| Gaming | MEDIUM | Same socket path as XR |
Gaps Surfaced¶
- gap-A: No native streaming/partial-hypothesis API — requires external wrapper.
- gap-B: No SSH/remote transport layer.
- gap-C: No code-vocabulary specialisation.