Transcribing streaming audio - Amazon Transcribe
Best practices

Transcribing streaming audio

Using Amazon Transcribe streaming, you can produce real-time transcriptions for your media content. Unlike batch transcriptions, which involve uploading media files, streaming media is delivered to Amazon Transcribe in real time. Amazon Transcribe then returns a transcript, also in real time.

Streaming can include pre-recorded media (movies, music, and podcasts) and real-time media (live news broadcasts). Common streaming use cases for Amazon Transcribe include live closed captioning for sporting events and real-time monitoring of call center audio.

Streaming content is delivered as a series of sequential data packets, or 'chunks,' that Amazon Transcribe transcribes instantaneously. The advantages of using streaming over batch include real-time speech-to-text capabilities in your applications and faster transcription times. However, this increased speed may have accuracy limitations in some cases.

Amazon Transcribe offers the following options for streaming:

To transcribe streaming audio in the AWS Management Console, speak into your computer microphone.

Tip

For SDK code examples, refer to the AWS Samples repository on GitHub.

Audio formats supported for streaming transcriptions are:

  • FLAC

  • OPUS-encoded audio in an Ogg container

  • PCM (only signed 16-bit little-endian audio formats, which does not include WAV)

Lossless formats (FLAC or PCM) are recommended.

Note

Streaming transcriptions are not supported with all languages. Refer to the 'Data input' column in the supported languages table for details.

To view the Amazon Transcribe Region availability for streaming transcriptions, see: Amazon Transcribe Endpoints and Quotas.

Best practices

The following recommendations improve streaming transcription efficiency:

  • If possible, use PCM-encoded audio.

  • Ensure that your stream is as close to real-time as possible.

  • Latency depends on the size of your audio chunks. If you're able to specify chunk size with your audio type (such as with PCM), set each chunk to between 50 ms and 200 ms. You can calculate the audio chunk size by the following formula: 

    chunk_size_in_bytes = chunk_duration_in_millisecond / 1000 * audio_sample_rate * 2

  • Use a uniform chunk size.

  • Make sure you correctly specify the number of audio channels.

  • With single-channel PCM audio, each sample consists of two bytes, so each chunk should consist of an even number of bytes.

  • With dual-channel PCM audio, each sample consists of four bytes, so each chunk should be a multiple of 4 bytes.

  • When your audio stream contains no speech, encode and send the same amount of silence. For example, silence for PCM is a stream of zero bytes.

  • Make sure you specify the correct sampling rate for your audio. If possible, record at a sampling rate of 16,000 Hz; this provides the best compromise between quality and data volume sent over the network. Note that most high-end microphones record at 44,100 Hz or 48,000 Hz.