Mediatavio
YouTube to MP3PriceBlogPrivacy

Transcription

AI Speech to Text: Language Models and Transcription Workflows

Choose a speech language, review the transcript and distinguish subtitle exports from document editing.

By Mediatavio Engineering

Make the language choice explicit

Speech to Text converts supported recordings into text. Load the recording, choose its spoken language and a model, run transcription, then review the words and timing before exporting TXT, SRT, VTT or JSON.

Automatic model selection is not automatic language detection. In this implementation, it routes the selected Chinese, English, Cantonese, Japanese or Korean language to SenseVoice Small; other selected languages route to Whisper, subject to the available model's support.

TXT is suitable for plain text. SRT and VTT contain timed segments. JSON preserves structured result information. The current export helpers do not provide PDF or DOCX, and a transcript is not a certified record.

How recognition and timestamps work

Audio is prepared as 16 kHz mono for the selected recognition model. A simple energy-based speech probe helps identify speech-bearing material; it is not a neural speaker diarization system.

Recognition applies options such as punctuation, beam size and hotwords and processes consecutive audio windows. Segment times are offset back to their positions in the whole recording. Chunk boundaries, pauses and model errors can affect both wording and timing.

The recognition call disables translation. This page transcribes the selected language rather than promising an English translation. It does not implement speaker identity labels or guarantee word-level alignment. Background music, overlapping speech and unusual names still require manual review.

Compare editing and processing workflows

Official transcription workflows; no word-error-rate benchmark
ServiceDocumented workflowBest-fit distinction
MediatavioWhisper/SenseVoice transcription; TXT, SRT, VTT and JSON exportsManual language selection, transcript review and timed subtitle exports
Adobe PodcastUpload, correct interactive transcripts and download PDF, DOCX or textHosted transcription connected to document-style audio/video editing
Azure batch transcriptionSubmit stored audio and retrieve asynchronous job resultsCloud API/service integration, not a like-for-like single-page browser utility

These distinctions come from the linked official documentation. No shared recording, word-error rate or timing test was performed. Language support, diarization options, costs and retention requirements must be checked for the exact service configuration.

Evaluate accuracy with a reference transcript

Use the same source and language, and manually transcribe a representative passage. Inspect substitutions, omitted words, invented words during silence, names and subtitle boundaries. Do not compare an automatically cleaned transcript with a raw result without disclosing the editing.

For a noisy source, compare transcription before and after noise cleanup; denoising can help or harm recognition.