文字起こしエージェントアクションの追加（Summer'26）
エージェントアクションにWhisper V3 Turbo （Whisper Large V3 Turboと思われる）による文字起こし機能が追加。6/22週から利用可能。
- Whisperはセルフホストっぽい。
- 5MB未満
- 話者識別やタイムスタンプはない。
電話音声だと帯域狭いから、適切な前処理すれば50分ぐらいいけるかも。MP3 64Kbpsだと10分ほどの計算になる。
Choose a Transcription Model for the Speech to Text Agent Action
Use the Transcription Model parameter on the Speech to Text action to control which model converts audio to text. Choose the model that best fits your transcription needs, such as audio quality, language, and data handling requirements. This parameter is available in Agentforce Builder, Flow, and REST.
Where: This change applies to Lightning Experience in Enterprise, Performance, Unlimited, and Developer Editions with Foundations or Agentforce 1 Editions.
When: This change is available starting the week of June 22, 2026.
How: In Agentforce Builder, open an agent that uses the Speech To Text action and expand the action's input parameters. Set the Transcription Model to Whisper V3 Turbo or ElevenLabs Scribe V2. If you don't set a value, Whisper V3 Turbo is used by default.
Considerations and Limitations for Speech to Text in Agentforce
Before using the Speech to Text action, review the considerations and limitations.
snip
- The Speech to Text agent action isn’t supported in the Government Cloud.
- The Speech to Text action converts supported audio files into text and returns the transcript as output.
- The Speech to Text action can be added to any subagent and any agent.
- This action is exposed through the invocable actions API and is available through agent actions, REST, Flow, and Apex.
- The Speech to Text action processes audio differently depending on the selected transcription model. Whisper V3 Turbo is hosted internally by Salesforce in your org's region, which can be relevant for compliance, and ElevenLabs Scribe V2 uses a third-party public API.
- Audio in unsupported or mixed languages might not be fully recognized.
- This action can be used in Flow and automation scenarios.
- This action is subject to invocation limits based on available memory in your organization. Requests that exceed the limit may fail.
- Processing time may vary based on audio length, file size, audio quality, and content complexity, as well as the transcription model selected. Whisper V3 Turbo is generally faster, while ElevenLabs Scribe V2 may have less predictable processing times.
- Error handling can be managed using Flow fault paths.
- Files larger than 5 MB aren't supported.
- Translation isn't supported.
- Real-time microphone capture isn't supported.
- Speaker identification, timestamps, and confidence scores aren't available.
- When adding the Speech to Text action to a subagent, ensure that the subagent instructions prompt for either a valid audio file ID or a file name. If a file name is provided, the subagent must include the Query Records action to resolve the file name to an audio file ID before invoking Speech to Text action.
Supported Audio Formats and File Limits
Item Supported Value
File Formats MP3, WAV, FLAC, OGG, OGA, AMR, MPEG, MPGA
Maximum File Size 5 MB
Input Type Audio File ID
Output Converted Text
Language Support
The Speech to Text action automatically detects the language of the audio.
Language detection of the audio is automatic and can’t be configured.
Mixed-language or unsupported-language audio may work, but transcription accuracy may vary.
Transcription quality may vary depending on the detected language. If you plan to use the transcript in Agentforce downstream features, verify that the detected language is supported. See Generative AI Supported Languages.
This action transcribes audio to text only and doesn’t perform translation. If the transcript will be used in Agentforce conversations or actions, verify that the transcript language is supported. See Generative AI Supported Languages.
Usage Types for Speech to Text
Digital Wallet Card usage type Description notes
Flex Credits Speech to Text Speech to Text converts audio input into text using Automatic Speech Recognition (ASR) models. Flex credit usage is metered based on the duration of audio processed, measured in seconds. Refer to Flex Credits Billable Usage Types
