Best local speech-to-text for subtitles and timestamps
Subtitle transcription needs more than readable text. You need stable segment boundaries, useful timestamps, punctuation, long-audio behavior and a workflow that can turn those timings into SRT or WebVTT without constant manual repair.
What makes an ASR stack good for subtitles?
- Timestamp quality: segment or word timing must remain aligned over long recordings.
- Segmentation: readable subtitle lines should follow phrases rather than arbitrary audio chunks.
- Punctuation and capitalization: raw words are not enough for publishable captions.
- Long-audio stability: silence, music and scene changes should not cause drift or hallucinated text.
- Speaker handling: interviews and meetings may need diarization on top of ASR.
- Language support: timing features are useless if the recognizer is weak or unsupported for the target language.
Parakeet v3: the cleanest model-first subtitle candidate
NVIDIA's Parakeet TDT 0.6B v3 publishes word-level and segment-level timestamps directly and also includes automatic punctuation and capitalization. Its official model card supports 25 languages and documents long-audio transcription workflows.
That combination makes Parakeet unusually well aligned with subtitle production. The hard gate is language coverage: if the target language is not in the official list, do not select it purely because the timestamp story is strong.
Whisper: the broadest subtitle ecosystem
Whisper remains attractive because there are many local runtimes and post-processing tools around it. The reference model is commonly used with segment timing, while faster-whisper exposes word timestamps and VAD options. If you need tighter word alignment or speaker labels, WhisperX adds forced phoneme alignment and optional diarization on top of a faster-whisper backend.
This modularity is both the advantage and the cost. A Whisper subtitle stack may involve ASR, VAD, forced alignment and diarization as separate components, which gives control but also increases dependencies and failure modes.
Qwen3-ASR: official forced alignment, but as a separate step
Qwen3-ASR supports long-audio transcription and provides a separate Qwen3-ForcedAligner-0.6B. The official project documents timestamp prediction for arbitrary units within up to five minutes of speech in 11 languages. That can produce a strong alignment workflow when your language is covered.
The important architectural detail is that recognition and precise alignment are separate. Also, Qwen's streaming inference path currently does not provide timestamps, so live captions and post-produced subtitles are not the same workflow.
VibeVoice-ASR: structured timestamps plus speaker attribution
VibeVoice-ASR is unusually relevant when subtitles come from interviews, meetings or podcasts with multiple speakers. Its official output format includes start time, end time, speaker ID and content in the same generated structure. For supported languages, that can collapse transcription, coarse timing and speaker attribution into one model pass.
The trade-off is weight: the current Hugging Face artifact reports 9B total parameters, so this is not the lightweight subtitle default. Also benchmark timestamp stability before using generated boundaries directly in an editor; a structured timestamp field is useful, but production subtitle timing still benefits from spot checks and, where needed, a dedicated aligner.
Where Moonshine and Nemotron fit
Moonshine and Nemotron 3.5 are more compelling when the main product requirement is low-latency live speech. They can still feed captioning applications, but they are not the first tools we would pick for a post-production subtitle pipeline where detailed timing, alignment and long-file editing dominate the workflow.
A practical subtitle benchmark
Do not evaluate only WER. Take 20–30 minutes of representative material and measure:
- word/segment timing drift near the end of the file;
- line breaks and phrase boundaries;
- names, numbers and punctuation;
- behavior across silence, music and overlapping speech;
- speaker attribution if needed;
- manual correction time per 10 minutes of finished subtitles.
The last metric is often the most important. A transcript with slightly better WER can still create more editing work if segmentation and timestamps are worse.
Recommended starting paths
| Need | First stack to test |
|---|---|
| Direct timestamps + punctuation in supported languages | Parakeet TDT 0.6B v3 |
| Broad runtime choice and customizable pipeline | Whisper + faster-whisper |
| Word alignment + optional speaker diarization | WhisperX |
| Qwen ASR with official detailed alignment | Qwen3-ASR + Qwen3 Forced Aligner |
Primary sources
Reviewed against primary sources on September 13, 2026. Model behavior, runtime support, language coverage and licenses can change; re-check the linked primary source before production use.