Features

Updated at:

Intelligent Speech Interaction provides speech recognition, speech synthesis, and speech analysis capabilities. This topic describes the features and use cases of each service to help you choose a service for your business needs.

Speech recognition

Speech recognition converts speech in audio into text. Choose a service based on the audio source, duration, and how quickly you need the results.

Service

Input method

Duration and file limits

Result delivery and processing time

Use cases

Short sentence recognition

Short audio

Up to one minute

Recognition of each submitted utterance

Voice search, short conversations, and voice commands

Real-time speech recognition

Continuous audio stream

Long, uninterrupted speech

Results returned as audio is sent

Live transcription of meetings, speeches, and live streams

Recording file recognition

HTTP URL of an existing file

Duration: ≤12 hours; audio: ≤512 MB; video: ≤2 GB

Asynchronous; within 24 hours for trials or 3 hours for the commercial edition

Non-real-time transcription of existing recordings or videos

Recording file recognition (flash version)

Direct audio file upload

Duration: ≤2 hours and file size: ≤100 MB

Synchronous; results in the same request

Short video editing, subtitle generation, and near-real-time quality inspection

Recording file recognition (idle version)

HTTP URL of an existing file

Duration: ≤12 hours; audio: ≤512 MB; video: ≤2 GB

Asynchronous; within 24 hours

Quality inspection, media management, and subtitles that can tolerate longer waits

The processing times for recording file recognition and the idle version do not apply to large batches of more than 500 hours of audio submitted within half an hour. Contact a presales specialist for these workloads. The flash and idle versions are available only in the commercial edition. Short sentence recognition, real-time speech recognition, and recording file recognition offer trial editions. For trial conditions, see Billing methods under Related documentation.

Note

For the full list of supported languages, dialects, and models, see the language tables in the API reference for each recognition service. Also check the audio sample rate when choosing a model. For example, telephone recordings typically use 8 kHz. Check language and dialect support for that sample rate.

Audio format, sample rate, and channel requirements also vary by service. See the corresponding API reference in the preceding table.

Speech synthesis

Speech synthesis converts text into speech. The console provides three service types: speech synthesis, long-text speech synthesis, and streaming text speech synthesis (CosyVoice). Choose a service based on text length, text input method, and audio delivery.

Service

Invocation mode

Text input method

Text length

Audio delivery

Use cases

Speech synthesis

Real-time synthesis

Single submission

≤300 characters

Streamed; playback can start as audio arrives

Short announcements and prompts

Long-text speech synthesis

Real-time synthesis

Single submission

≤10,000 characters

Streamed; playback can start as audio arrives

Reading articles or novels aloud with prompt playback

Asynchronous synthesis

Single submission

≤100,000 characters

Asynchronous; completes within 3 hours

Generating audio for long articles or novels in advance

Streaming text speech synthesis (CosyVoice)

Streaming synthesis

Multiple batches per session

See the API reference for batch limits

Streamed; no need to wait for the complete text

Speech output for streaming text from a large language model

To read prepared long text with CosyVoice voices, you can also use the CosyVoice long-text speech synthesis API. This API accepts a single text submission of up to 10,000 characters per session and streams the resulting audio. Streaming text input is not supported.

Speech synthesis offers trial and commercial editions. Long-text speech synthesis and both CosyVoice synthesis APIs described above support only the commercial edition and do not offer trials. For speech synthesis, real-time long-text synthesis, and asynchronous long-text synthesis, each Chinese character, English letter, punctuation mark, or space within a sentence counts as one character. Check text length limits and billable character counts separately for each service; these counting rules are not interchangeable.

Note

For supported voices, languages, and audio formats, see the voice lists and parameter descriptions in the corresponding service documentation.

On-device speech synthesis

Offline speech synthesis uses a local model on the device to convert text into speech without a network connection during synthesis. It is suitable for offline announcements, audio reading, and accessibility announcements. The SDK must be activated before use. Charges are based on the number of activated devices.

Speech analysis

Speech analysis detects sound events in audio or analyzes speakers' voice characteristics and languages.

Service

Analysis target

Output

Use cases

Sound event detection

Sound events such as music, crying, laughter, and explosions

Sound events with start and end times

Safety monitoring, content moderation, and assisted audio/video editing

Speaker verification

Input speech and a specified user's voice characteristics

1:1 verification result

Voice-based login and customer identity verification

Gender identification

The speaker's voice

Gender identification result

User profiling

Language identification

Chinese, English, or Cantonese speech

Language identification result, not a transcription

Language identification in customer service, meetings, and voice commands

Related documentation

  • For differences between the trial and commercial editions, trial quotas, and upgrade instructions, see Billing methods.

  • For billing rules and resource plans for each service, see Billable items.

  • For concurrency and QPS limits, see Concurrency and QPS.