Is my recording uploaded?
No. The recording bytes are processed locally. Your browser separately downloads the AI runtime/model files needed for transcription.
AI Transcriber & Subtitle Generator
Enable JavaScript to transcribe media locally.
About this tool
Private speech-to-text in your browser
NowConverter uses browser-compatible OpenAI Whisper models through Transformers.js. The AI runtime and model files are fetched when needed, then inference runs locally with WebGPU when available or a background CPU/WASM compatibility worker. Your selected recording is decoded and transcribed on your device rather than uploaded to a speech-recognition server.
What you can export
Review the editable transcript and timestamped subtitle cues, then copy the text or download TXT, SRT, VTT or JSON. Clicking a subtitle cue jumps the original media player to that timestamp so you can quickly verify wording and timing.
Accuracy and browser limits
Whisper is strong on clear speech and many languages, but no speech recognizer is perfect. Background noise, overlapping speakers, strong reverberation, unusual names and low-volume recordings can reduce accuracy. Larger models can improve recognition but also require bigger downloads, more memory and longer processing time. Tamil and some lower-resource languages are harder for the general Whisper models; manually selecting the language and using Accurate mode on WebGPU can improve results, but review is still important. Original-language transcription now requires a manual language choice so Whisper does not unexpectedly return English. Video formats the browser cannot decode directly may use NowConverter's local FFmpeg fallback to extract the audio track.
Can it auto-detect language?
Auto-detect is available for Translate to English. For original-language transcription, NowConverter asks you to choose the spoken language manually because browser Whisper can sometimes return English when language detection is uncertain.
Can it translate speech into English?
Yes. Choose “Translate speech to English.” The multilingual Whisper model performs the translation locally while transcribing.
Why is Tamil sometimes less accurate?
Whisper performs unevenly across languages. For Tamil, manually choose Tamil instead of Auto-detect and use Accurate mode with Auto/WebGPU when available. NowConverter also applies a slower multilingual decoding boost for selected harder languages.
Why can the first run take longer?
The selected Whisper model must be downloaded before it can run. Browsers may cache model files, so later sessions can start faster depending on cache settings and available storage.
Treat transcription as a draft that still needs human review
Speech recognition is sensitive to accents, background noise, overlapping speakers, names and recording quality. A transcript can look fluent while still containing an important wrong word, number or proper noun. For interviews, subtitles, legal notes or research, listen back to uncertain passages instead of assuming grammatical text is accurate.
Subtitle formats add timing information as well as words. SRT is widely supported and intentionally simple. WebVTT adds web-oriented features and is the usual choice for HTML5 video. If the final subtitles will be published, review line length, reading speed and timing after transcription rather than treating the raw AI segmentation as the finished edit.
Practical checks
- Choose the spoken language manually when accurate original-language output matters.
- Proofread names, dates, measurements and numbers against the recording.
- Use SRT for broad compatibility and VTT for many web-video workflows.
Practical walkthrough · ai transcriber
Turn a recorded interview into editable subtitles
Example task: An interview has clear dialogue and needs a first-pass SRT for video editing.
Try this workflow
- Choose the correct spoken language and an appropriate processing mode.
- Generate the transcript locally, then listen back to uncertain sections.
- Correct names and timing, export SRT, and preview the subtitles on the actual video.
What to expect
SRT contains timed caption cues. The transcript is a draft that still needs listening-based review.
When to take extra care
Model downloads, GPU/CPU compatibility and language quality vary; fast speech, crosstalk and Tamil words may require extensive corrections.