> For the full page index, fetch https://echonova.ai/llms.txt # Speech to Text formats Which files the Speech to Text page accepts, what the browser does to each format before upload, what a microphone take is saved as, and what the transcript is. This page is the lookup table for what goes into the **Speech to Text** page and what comes out: the seven file formats it accepts, what your browser does to each one before it is sent, what a microphone take is saved as, the checks that run once the file has been sent, and what the transcript contains. For the limits and the credit rate, see [Speech to Text limits and credits](/docs/playground/stt-limits-and-credits). ## Accepted files The page accepts a file by its extension. The file picker filters to these types; a file dropped on the drop area is checked the same way. | Format | Extension | Before upload | | ------ | --------- | ------------------------------- | | MP3 | `.mp3` | Converted to WAV by the browser | | WAV | `.wav` | Sent unchanged | | FLAC | `.flac` | Sent unchanged | | OGG | `.ogg` | Sent unchanged | | M4A | `.m4a` | Converted to WAV by the browser | | AAC | `.aac` | Converted to WAV by the browser | | WEBM | `.webm` | Converted to WAV by the browser | Any other extension is refused before upload with **Please upload one of: MP3, WAV, FLAC, OGG, M4A, AAC, WEBM.** The badges under the drop area name four of the seven; all seven are accepted. ## What the browser does before upload A WAV, FLAC or OGG file is sent as you chose it. Every other accepted format is prepared in the browser first: 1. The browser decodes the file. 2. Stereo is mixed down to one channel; both channels are averaged, so a call recording with one quiet side keeps both voices. 3. The audio is resampled to 16 kHz. 4. The result is written as a 16-bit WAV, named after your file with a `.wav` extension, and that copy is sent. What this means for you: - **Playback uses the converted copy.** The player on the transcript card, the play buttons in the **Transcripts** list and the audio kept in **Assets** all carry the 16 kHz mono WAV, not the file you chose. The name shown on the transcript card is still your original file name. - **The 150 MB limit is checked twice**: on your original file before conversion, and on the converted copy after it. One hour of 16 kHz mono WAV is about 115 MB, so a converted file inside the 1-hour limit always passes the second. - **Preparation takes time.** The status line shows **Preparing audio…** while the browser works. A long MP3 can hold there for a while before **Uploading** and its percentage appear. - **A file the browser cannot decode is sent unchanged.** The transcription then ends with a reason that starts with **Audio could not be decoded** and the charge is refunded. Export the recording again from the application that made it. ## Microphone takes **Record** captures the take in the page itself, as 16 kHz mono 16-bit WAV, whichever browser you use. It needs a browser with microphone capture and audio worklet support; otherwise the page says **Your browser does not support audio recording.** | What you record in | Recorded as | Extension of the take | Before upload | | --------------------- | --------------- | --------------------- | -------------- | | Any supported browser | 16 kHz mono WAV | `.wav` | Sent unchanged | The take is named `recording-` followed by a number, and it goes through the same checks as an uploaded file. A take stops on its own at 1 hour. Press **Transcribe** to send it. ## Checks after upload Once the file has been sent, more checks run before it is accepted and charged. Each refusal shows in the status line under the box, followed by a short code in square brackets, with a **Try again** button. | Check | Rule | Message | | --------- | ---------------------------------------------- | ---------------------------------------------------------------------------------------- | | File type | The file must be audio | **Unsupported file type: … Please upload an audio file.**, naming the type that was sent | | Size | 150 MB or less | **File too large** | | Length | 1 hour or less, measured by decoding the audio | **Audio exceeds the 1-hour limit.** | | Readable | The audio length can be measured | **Could not read the audio length. Try another file.** | The server measures the length by decoding the audio, not from the file's header, and the credit charge uses that measured length. See [Speech to Text limits and credits](/docs/playground/stt-limits-and-credits). After acceptance, the audio is decoded. WAV, FLAC and OGG with Opus decode; a stereo file is averaged to mono; audio shorter than 50 milliseconds ends with a reason that starts with **Audio is too short**. Every failure after acceptance refunds the charge. ## The transcript | Fact | Value | | ----------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Language | Auto, Arabic or English, picked on the page. Auto lets the service decide, which is Arabic today. | | Shape | With **Identify speakers** on, the default: rows in the order they were said, each with its start time and its speaker. With it off: one block of text per file. No word timestamps and no confidence figures. | | Shown on the page | A transcript card with the file name, its length and the language, and three tabs: **Text** (the rows or the block), **Subtitles (SRT)** and **JSON** (the files themselves). | | Saved as | A TXT, SRT or JSON file, matching the open tab, named after your file with its extension replaced, from **Download**; **Copy** copies the open tab instead. SRT holds cues of at most 7 seconds, each starting with its speaker; a transcript in one block gives one cue for the whole file. | How to read the result view, and where the transcript is kept afterwards, is on [Read a transcript](/docs/playground/reading-results). ## Worked example You drop `memo.m4a`, a 3-minute voice memo from your phone, on the drop area. The extension passes, the browser reads its length as 3:00, and you press **Transcribe**. The status line shows **Preparing audio…** while the memo is decoded, mixed to mono, resampled to 16 kHz and written as `memo.wav`, about 5.8 MB. It then shows **Uploading** with a percentage and switches to **Transcribing**. When the transcript arrives, the transcript card is titled `memo.m4a`, the player plays the WAV copy, and **Download** on the **Text** tab saves `memo.txt`. ## FAQ ### My file is an MP3. Will it work? Yes. The browser converts it to WAV before upload. If the conversion fails, the transcription reports it and the charge is refunded. ### Why is the downloaded audio a WAV when I uploaded an M4A? The page transcribes the converted copy and keeps that copy. Keep your original if you need it in its own format. ### Why is my one-hour WAV refused when a one-hour MP3 is accepted? The 150 MB limit applies to the file as you chose it. A one-hour MP3 is about 57 MB; a one-hour stereo 44.1 kHz WAV is far larger. Export the WAV as MP3, or as 16 kHz mono, before you upload it. ### Can I get timestamps or speaker labels? Yes. With **Identify speakers** on, which is the default, each row of the transcript shows its start time and its speaker, and the **Subtitles (SRT)** tab and its download carry both. There are no timestamps for single words. See [Read a transcript](/docs/playground/reading-results). ### Does a microphone take sound different after conversion? The page records the take at 16 kHz mono, which is what transcription uses, so no conversion happens afterwards. The transcript is unaffected. ## Related - [Transcribe a file](/docs/playground/transcribe-a-file): the steps from upload to transcript. - [Read a transcript](/docs/playground/reading-results): every control on the result view. - [Speech to Text limits and credits](/docs/playground/stt-limits-and-credits): every limit and the credit rate. - [Speech to Text formats and output](/docs/developers/stt-formats-and-output): the same checks as the API applies them.