AI Audio Transcriber — Free Whisper, 99 Languages

The tool runs OpenAI Whisper on the Hugging Face Inference API. The default checkpoint is openai/whisper-large-v3 (1550 M parameters, the lowest published WER of the public Whisper checkpoints); a faster whisper-medium (769 M) is offered for English-only dictation. We keep transcription on the server because every published quantised Whisper is at least 75 MB — too much to ship into a Sri Lankan mobile browser — and the larger checkpoints, the ones that actually read Sinhala and Tamil well, are out of reach for in-browser delivery entirely.

A single transcription goes through five deterministic steps:

1. Validate. The file must be an audio or video format ffmpeg can decode (MP3, WAV, M4A/AAC, OGG, FLAC, MP4, WebM, MOV) and at most 25.0 MB.
2. Forward. The bytes are POSTed once to /api/tools/transcribe-audio, which streams them to api-inference.huggingface.co with the chosen language and task. Nothing is persisted on our server — the bytes live only in request memory and are released when the response returns.
3. Decode + resample. The HF backend uses ffmpeg to decode the container and resample to 16,000 Hz mono — the sample rate Whisper's feature extractor expects.
4. Chunk + decode. Whisper natively processes 30-second windows. The HF pipeline is called with return_timestamps: true and chunk_length_s: 30 so cues stay aligned across chunk boundaries on files longer than 30 seconds. Decoding is greedy (no sampling) with the source language token forced if you picked one, or Whisper's built-in language-detection token if you left the picker on Auto-detect. The task token selects transcribe or translate (any-language → English).
5. Build cues. The API returns { text, chunks: [{ text, timestamp:[start,end] }, …] }. SRT cues are built with HH:MM:SS,mmm timestamps (SubRip convention — comma separator). WebVTT cues use the same timestamps with a dot instead of a comma, per the W3C WebVTT specification. A second independent formatter (formatTimestampByMath) cross-checks the primary formatter against pure-arithmetic arithmetic on every cue — a build-time assert flags any drift.

End-to-end latency depends on the HF queue and the chosen checkpoint. A 1-minute clip on whisper-large-v3 typically returns in 15–30 seconds; a 5-minute clip in 30–90 seconds. Medium is about twice as fast and is the right choice for English-only speech where the slight accuracy hit is invisible. The server enforces a 120-second timeout per request and gracefully reports a friendly “model unavailable” banner when the upstream is queued or otherwise unreachable, so the page never shows a dead state.

Two privacy notes. The audio bytes are forwarded as a singlemultipart/form-data POST and discarded as soon as the upstream response returns — there is no log of audio content, only a server-side timing counter (request ms, model id). And the tool never attempts speaker identification, age, or ethnicity inference — Whisper is a sequence-to-sequence ASR model, not a speaker classifier.

Whisper large-v3 is the default — best Sinhala and Tamil quality of any public Whisper checkpoint. Medium is about twice as fast on the HF infrastructure and is the right pick for English-only dictation where the per-language WER gap is small. The WER numbers below are the multilingual test scores reported in the Whisper paper.

• Radford et al. — Robust Speech Recognition via Large-Scale Weak Supervision (Whisper paper, 2022)
• Hugging Face — openai/whisper-large-v3 model card
• Hugging Face — openai/whisper-medium model card
• Hugging Face — Inference API reference (automatic-speech-recognition)
• W3C — WebVTT 1.0 specification (caption file format)
• SubRip — SRT file format reference

The Whisper paper, model cards, Hugging Face Inference API reference, and the WebVTT specification were all cross-checked on 2026-05-12. The page is reviewed quarterly and whenever Hugging Face rotates the Whisper variants in its hosted inference catalogue.