Speech recognition
Speech recognition is speech-to-text (STT) conversion.
To work with SpeechKit, you can access it via the API or Playground. For more information about working with the Yandex Cloud API, see API concepts.
The service is available at stt.api.cloud.yandex.net:443.
You can also work with SpeechKit using the Python SDK. It is implemented based on the SpeechKit API v3.
To try out the Text-to-Speech and Speech-to-Text product demos, visit the SpeechKit page on our website.
Recognition methods
SpeechKit provides two ways of improving the quality of speech recognition:
- Streaming recognition is used for real-time speech recognition. During streaming recognition, SpeechKit receives short audio fragments and sends the results, including intermediate ones, over a single connection.
- Audio file recognition. SpeechKit Can recognize audio recordings in synchronous and asynchronous mode.
- Synchronous mode has strict limitations on the size and duration of a file and is suitable for recognizing single-channel audio fragments of up to 30 seconds.
- Asynchronous mode can process multi-channel audio fragments. Maximum recording duration: 4 hours.
Which recognition to choose
| Streaming recognition | Synchronous recognition | Asynchronous recognition | |
|---|---|---|---|
| Use cases | Telephone assistants and robots Virtual assistants | Virtual assistants Voice control Recognition of short voice messages in messengers | Transcription of audio calls and presentations Subtitling Ensuring script adherence in call centers Identifying successful scripts Evaluating performance of call center operators. |
| Input data | Real-time voice | Pre-recorded short single-channel audio files | Pre-recorded multi-channel and long audio files |
| How it works | Exchanging messages with the server over a single connection | Request — quick response | Request — delayed response |
| Supported APIs | gRPC v2 gRPC v3 | REST v1 | REST v2 gRPC v3 |
| Maximum duration of audio data | 5 minutes | 30 seconds | 4 hours |
| Maximum amount of transmitted data | 10 MB | 1 MB | 1 GB |
| Number of recognition channels | 1 | 1 | 2 |
Recognition process
Audio is recognized in three stages:
- The acoustic model determines which set of low-level attributes corresponds to the audio signal.
- The language model uses the acoustic model output to generate the text by words.
- The service performs text processing: punctuation, converting numerals into numbers, and more.
Recognition accuracy
Recognition accuracy depends on the recognition model. You can improve recognition accuracy of the model by providing data for tuning of the model. For more information about model tuning, see Extending a speech recognition model.
The accuracy of speech recognition is also affected by:
- Original sound quality.
- Audio encoding quality.
- Speech intelligibility and rate.
- Utterance complexity and length.