Speech to text Turning a recorded meeting into words, and how far to trust the result.
Short answer. Speech to text, also called automatic speech recognition (ASR), turns spoken audio into written words. Jurafsky and Martin define the task as mapping an acoustic waveform to a string of words. Accuracy depends on conditions: their textbook puts error rates around 2% on clean English read speech, 5.8 to 11% on conversations such as AMI meetings, and up to 25.5% on distant-microphone dinner-party speech. Open models such as Whisper can run on an office’s own hardware, so recordings need not leave the building, but accents, noisy rooms and Hindi-English mixing still raise errors, so test on your own meetings.
01
What happens to the audio
The audio is sampled and converted to a spectral picture of the sound. In OpenAI’s Whisper model the audio is resampled to 16,000 Hz and turned into an 80-channel log-Mel spectrogram on 25-millisecond windows with a 10-millisecond stride. A neural network then predicts the words, with punctuation, and usually timestamps.
The result is text, not minutes and not a list of who said what. Splitting speakers is a separate task called speaker diarization, and turning a transcript into decisions is a later step.
02
Why conditions matter
Jurafsky and Martin list four things that make recognition harder: a large vocabulary, people talking to each other rather than to a machine, noise or a distant microphone, and accents unlike the training data. They single out conversational speech, such as transcribing a business meeting, as the hardest case on the second of those.
Condition
Typical word error rate
Source
English read speech, audiobooks (LibriSpeech clean)
About 2%
Jurafsky and Martin
Conversations: AMI meetings, Switchboard and CALLHOME phone calls
5.8 to 11%
Jurafsky and Martin
Four people at a dinner party, distant microphones (CHiME-6)
Up to 25.5%
Jurafsky and Martin
English read sentences, Whisper large-v2 on FLEURS
4.2%
Whisper paper
Hindi read sentences, Whisper large-v2 on FLEURS
21.5%
Whisper paper
03
Batch or live
The textbook notes that attention-based models reach higher accuracy, while models trained with the CTC loss adapt more easily to streaming, where words appear while the person is still speaking. Whisper’s model card says it cannot be used for real-time transcription out of the box, although other projects build near-real-time tools on top of it.
Mode
How it works
Use in an office
Batch (a recorded file)
The whole recording is processed, in 30-second windows for Whisper-style models
Meeting minutes, interviews, hearings
Streaming (live)
Words appear as the person speaks; CTC-style models suit it
Live captions, voice agents, dictation
04
Office checklist
Get consent before recording: Whisper’s model card cautions against transcribing recordings of individuals taken without their consent.
Put a microphone near each speaker where you can; distance makes recognition harder.
Decide the language mix in advance, and see our pages on Hinglish and code-switching.
Test a ten-minute stretch of a real meeting and count the errors; our word error rate page shows how.
Check names, numbers and dates by hand, and keep a person in the loop for anything that will be signed.
05
Where ANVE stands
ANVE’s minutes maker produces transcripts in Hindi, English and Hinglish on a device in the office. Every action item must quote the transcript line it came from, and a person confirms each owner before it is approved. The sealed unit, which is in design, adds a physical switch with a light that cuts the microphones. Whichever product you choose, ANVE’s included, test it on your own recordings.
06
Questions people ask
Is speech to text the same as dictation?
Dictation is one use of it, with one speaker close to a microphone. Jurafsky and Martin note that people speaking to machines are easier to recognise than people speaking to each other, and that conversational speech such as a business meeting is the hardest.
Can speech to text work offline?
Yes. Whisper’s code and weights are openly released, so they can run on a local machine. In the faster-whisper project’s test, a GPU transcribed 13 minutes of audio with the large-v2 model in about a minute.
How accurate is speech to text in Hindi?
In OpenAI’s 2022 paper, Whisper large-v2 scored a 21.5% word error rate on Hindi read sentences in the FLEURS test, against 4.2% for English. Meeting speech and Hindi-English mixing are harder, so test your own recordings.
Does it tell me who is speaking?
Not by itself. Working out who spoke when is a separate task called speaker diarization.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
