Microsoft AI MAI-Transcribe 2 is the recorded-audio speech-to-text edition of Microsoft’s second-generation transcription family. It converts spoken content into written text across 60 languages, with automatic language identification and support for speech that switches between languages. The model adds speaker diarization, word-level timing and selectable transcription styles. Useful artifacts include an attributed meeting transcript, a searchable interview record, caption-alignment data or checked dictation containing specialist terminology.
Prepare the recording with the spoken material intact. The maker lists WAV, MP3 and FLAC input. Keep the original audio available for comparing uncertain words, names and numbers with the recognized text. A recording can contain background noise or variable microphone quality, which the model is designed to handle, but the review should still concentrate on unclear or overlapping passages. For an interview, keep the sequence of turns and pauses rather than removing the context that helps interpret a response.
Automatic language detection is useful when the language is not known beforehand. The documented 60-language set includes English, Russian, Chinese, Japanese, Korean, Arabic, French, German, Spanish, Indonesian, Malay, Thai and Vietnamese. Code switching covers natural changes between languages, including blended pairs such as Hinglish or Spanglish. The maker’s interface also accepts a single language hint, which it describes as a strong instruction. Use that only when the recording is actually in the specified language rather than forcing a multilingual conversation into one expected language.
Keyword biasing supplies domain-specific terms, abbreviations, product names and proper nouns that may otherwise be difficult to distinguish from the sound. Prepare a focused list related to the recording, such as the project name, a product line or technical words used in the discussion. The serving guide describes these as recognition hints rather than forced matches. For a technical lecture, include the exact spellings of named instruments and specialist terms, then check whether the transcript places them where they were actually spoken. A supplied term is not evidence that the speaker said it.
Choose the output style from the intended artifact. Verbatim, the documented default, retains fillers, false starts and self-corrections. Clean removes disfluencies for a more readable transcript. A record used to examine how a conversation unfolded can preserve the hesitation and correction; a readable lesson transcript can use the cleaned wording. Keep the original audio alongside either result. Cleaning improves presentation of the spoken text and should not be treated as a separate summary or a new interpretation of what the speaker meant.
Speaker diarization separates turns by detected speaker, and word-level timestamps supply an offset and duration for each word. The serving interface also offers segment timing or no timing. Segments are divided by language and, when diarization is enabled, by speaker. Timing fields use milliseconds and can support navigation, search and editing. The labels distinguish detected speakers; matching a label to a participant’s name is a separate review step. For captions, use checked word timings to prepare readable cue groups rather than assuming the recognition result is already a finished subtitle file.
The model card lists a maximum file size of 300 MB and two hours. The linked synchronous REST reference for the documented 2025-10-15 interface instead requires a file smaller than 250 MB and shorter than two hours. Those are distinct published model-card and serving constraints; prepare a file within the applicable serving allowance. The MAI guide also states that diarization supports shorter recordings than transcription without diarization. Organize long multi-speaker material into coherent sections when speaker attribution is needed, and retain the section’s position in the original recording for later review.
The documented output is structured transcription data: the full text, phrases and, when requested, speaker and timing information. A phrase can carry its text, language, offset and duration; word timing can link a recognized word to the audio. For a searchable archive, preserve those relationships so a found term can lead back to the recording. For a meeting workflow, a checked transcript becomes an input to later summarization or action-item extraction. The speech model’s result is the recognized content, not an independently verified account of the meeting’s decisions.
The version distinction is concrete. MAI-Transcribe 1.5 covers 43 languages, while this edition covers 60 and adds the documented diarization, word-timestamp and clean/verbatim controls. MAI-Transcribe 2 Streaming is a separate model for incremental real-time transcription with its own feature set; its first-partial latency does not describe this recorded-audio edition. Check the transcript against the audio for names, numbers, speaker turns and mixed-language passages before using it as the source for captions, quotes or a derived document.
", "aboutText": "Microsoft AI MAI-Transcribe 2 is the recorded-audio speech-to-text edition of Microsoft’s second-generation transcription family. It converts spoken content into written text across 60 languages, with automatic language identification and support for speech that switches between languages. The model adds speaker diarization, word-level timing and selectable transcription styles. Useful artifacts include an attributed meeting transcript, a searchable interview record, caption-alignment data or checked dictation containing specialist terminology.\n\nPrepare the recording with the spoken material intact. The maker lists WAV, MP3 and FLAC input. Keep the original audio available for comparing uncertain words, names and numbers with the recognized text. A recording can contain background noise or variable microphone quality, which the model is designed to handle, but the review should still concentrate on unclear or overlapping passages. For an interview, keep the sequence of turns and pauses rather than removing the context that helps interpret a response.\n\nAutomatic language detection is useful when the language is not known beforehand. The documented 60-language set includes English, Russian, Chinese, Japanese, Korean, Arabic, French, German, Spanish, Indonesian, Malay, Thai and Vietnamese. Code switching covers natural changes between languages, including blended pairs such as Hinglish or Spanglish. The maker’s interface also accepts a single language hint, which it describes as a strong instruction. Use that only when the recording is actually in the specified language rather than forcing a multilingual conversation into one expected language.\n\nKeyword biasing supplies domain-specific terms, abbreviations, product names and proper nouns that may otherwise be difficult to distinguish from the sound. Prepare a focused list related to the recording, such as the project name, a product line or technical words used in the discussion. The serving guide describes these as recognition hints rather than forced matches. For a technical lecture, include the exact spellings of named instruments and specialist terms, then check whether the transcript places them where they were actually spoken. A supplied term is not evidence that the speaker said it.\n\nChoose the output style from the intended artifact. Verbatim, the documented default, retains fillers, false starts and self-corrections. Clean removes disfluencies for a more readable transcript. A record used to examine how a conversation unfolded can preserve the hesitation and correction; a readable lesson transcript can use the cleaned wording. Keep the original audio alongside either result. Cleaning improves presentation of the spoken text and should not be treated as a separate summary or a new interpretation of what the speaker meant.\n\nSpeaker diarization separates turns by detected speaker, and word-level timestamps supply an offset and duration for each word. The serving interface also offers segment timing or no timing. Segments are divided by language and, when diarization is enabled, by speaker. Timing fields use milliseconds and can support navigation, search and editing. The labels distinguish detected speakers; matching a label to a participant’s name is a separate review step. For captions, use checked word timings to prepare readable cue groups rather than assuming the recognition result is already a finished subtitle file.\n\nThe model card lists a maximum file size of 300 MB and two hours. The linked synchronous REST reference for the documented 2025-10-15 interface instead requires a file smaller than 250 MB and shorter than two hours. Those are distinct published model-card and serving constraints; prepare a file within the applicable serving allowance. The MAI guide also states that diarization supports shorter recordings than transcription without diarization. Organize long multi-speaker material into coherent sections when speaker attribution is needed, and retain the section’s position in the original recording for later review.\n\nThe documented output is structured transcription data: the full text, phrases and, when requested, speaker and timing information. A phrase can carry its text, language, offset and duration; word timing can link a recognized word to the audio. For a searchable archive, preserve those relationships so a found term can lead back to the recording. For a meeting workflow, a checked transcript becomes an input to later summarization or action-item extraction. The speech model’s result is the recognized content, not an independently verified account of the meeting’s decisions.\n\nThe version distinction is concrete. MAI-Transcribe 1.5 covers 43 languages, while this edition covers 60 and adds the documented diarization, word-timestamp and clean/verbatim controls. MAI-Transcribe 2 Streaming is a separate model for incremental real-time transcription with its own feature set; its first-partial latency does not describe this recorded-audio edition. Check the transcript against the audio for names, numbers, speaker turns and mixed-language passages before using it as the source for captions, quotes or a derived document.", "editorialOverlay": { "copy_fields": null, "copy_sha256": null, "copy_source": null, "html_sha256": "e7d73eac1ab41080856da6fbb7ddb7e13e0455809f596f6fdfc0f02d0c81cc8d", "module_sha256": "ee9dc1c50b1e6a8fb7fa062030fcb4646b9aaaeb08be513b723a5a92a7290f06", "raw_model_unchanged": true, "schema": "ailoot-native-model-editorial-v1" }, "locale": "en", "methods": [ { "html": "microsoft/mai-transcribe-2Microsoft AI MAI-Transcribe 2 turns recorded speech into text across 60 languages, with automatic language identification and code switching. It adds speaker diarization, word timing, keyword biasing and clean or verbatim styles for meeting records, captions, searchable audio and specialist dictation.
This description covers the model’s capabilities. Available in Ailoot inputs, outputs and options are listed for the selected method below.
Microsoft AI MAI-Transcribe 2 is the recorded-audio speech-to-text edition of Microsoft’s second-generation transcription family. It converts spoken content into written text across 60 languages, with automatic language identification and support for speech that switches between languages. The model adds speaker diarization, word-level timing and selectable transcription styles. Useful artifacts include an attributed meeting transcript, a searchable interview record, caption-alignment data or checked dictation containing specialist terminology.
Prepare the recording with the spoken material intact. The maker lists WAV, MP3 and FLAC input. Keep the original audio available for comparing uncertain words, names and numbers with the recognized text. A recording can contain background noise or variable microphone quality, which the model is designed to handle, but the review should still concentrate on unclear or overlapping passages. For an interview, keep the sequence of turns and pauses rather than removing the context that helps interpret a response.
Automatic language detection is useful when the language is not known beforehand. The documented 60-language set includes English, Russian, Chinese, Japanese, Korean, Arabic, French, German, Spanish, Indonesian, Malay, Thai and Vietnamese. Code switching covers natural changes between languages, including blended pairs such as Hinglish or Spanglish. The maker’s interface also accepts a single language hint, which it describes as a strong instruction. Use that only when the recording is actually in the specified language rather than forcing a multilingual conversation into one expected language.
Keyword biasing supplies domain-specific terms, abbreviations, product names and proper nouns that may otherwise be difficult to distinguish from the sound. Prepare a focused list related to the recording, such as the project name, a product line or technical words used in the discussion. The serving guide describes these as recognition hints rather than forced matches. For a technical lecture, include the exact spellings of named instruments and specialist terms, then check whether the transcript places them where they were actually spoken. A supplied term is not evidence that the speaker said it.
Choose the output style from the intended artifact. Verbatim, the documented default, retains fillers, false starts and self-corrections. Clean removes disfluencies for a more readable transcript. A record used to examine how a conversation unfolded can preserve the hesitation and correction; a readable lesson transcript can use the cleaned wording. Keep the original audio alongside either result. Cleaning improves presentation of the spoken text and should not be treated as a separate summary or a new interpretation of what the speaker meant.
Speaker diarization separates turns by detected speaker, and word-level timestamps supply an offset and duration for each word. The serving interface also offers segment timing or no timing. Segments are divided by language and, when diarization is enabled, by speaker. Timing fields use milliseconds and can support navigation, search and editing. The labels distinguish detected speakers; matching a label to a participant’s name is a separate review step. For captions, use checked word timings to prepare readable cue groups rather than assuming the recognition result is already a finished subtitle file.
The model card lists a maximum file size of 300 MB and two hours. The linked synchronous REST reference for the documented 2025-10-15 interface instead requires a file smaller than 250 MB and shorter than two hours. Those are distinct published model-card and serving constraints; prepare a file within the applicable serving allowance. The MAI guide also states that diarization supports shorter recordings than transcription without diarization. Organize long multi-speaker material into coherent sections when speaker attribution is needed, and retain the section’s position in the original recording for later review.
The documented output is structured transcription data: the full text, phrases and, when requested, speaker and timing information. A phrase can carry its text, language, offset and duration; word timing can link a recognized word to the audio. For a searchable archive, preserve those relationships so a found term can lead back to the recording. For a meeting workflow, a checked transcript becomes an input to later summarization or action-item extraction. The speech model’s result is the recognized content, not an independently verified account of the meeting’s decisions.
The version distinction is concrete. MAI-Transcribe 1.5 covers 43 languages, while this edition covers 60 and adds the documented diarization, word-timestamp and clean/verbatim controls. MAI-Transcribe 2 Streaming is a separate model for incremental real-time transcription with its own feature set; its first-partial latency does not describe this recorded-audio edition. Check the transcript against the audio for names, numbers, speaker turns and mixed-language passages before using it as the source for captions, quotes or a derived document.
POST /v1/audio/transcriptionsaudiotext| Field | Type | Requirement | Values & limits | Default |
|---|---|---|---|---|
Fieldmodel | Typestring | RequirementRequired | Values & limits microsoft/mai-transcribe-2 | Default— |
Fieldinput_audio | Typeobject | RequirementRequired | Values & limits | Default— |
Fieldlanguage | Typestring | RequirementOptional | Values & limits | Default— |
Fieldtemperature | Typenumber | RequirementOptional | Values & limits Minimum: 0 Maximum: 1 | Default— |
Fieldresponse_format | Typestring | RequirementOptional | Values & limits json, verbose_json | Defaultjson |
Fieldtimestamp_granularities | Typearray | RequirementOptional | Values & limits | Default— |
Fieldprovider | Typeobject | RequirementOptional | Values & limits | Default— |
| Parameter | Contract |
|---|---|
Parametercontent_type | Contractapplication/json |
Parameterrequired | Contractmodel, input_audio |
Parametermaterial_types | Contractaudio |
Parametermaterial_conditions | ContractRead an actual WAV file and encode its real bytes for the declared JSON input_audio; no multipart route is advertised for this exact JSON-only declaration. |
Lifecycle: synchronous_result_then_owned_library
GET /v1/media/tasks/{task_id} · actual returned task_id or media library record
GET /v1/media/assets/{asset_id}/content
GET /v1/media/assets
GET /v1/media/assets/{asset_id}
The displayed rates estimate the cost before generation. The final charge is calculated from the cost recorded for the original generation. Measured output duration and forecast quantities are not the final billable quantity.
Tariff references do not add request options. Supported parameters are listed for the selected method.
Match each charge to the conditions shown. Use one rate for each set of alternative configurations.
| Configuration | Rate schedule | Forecast quantity | Billing unit | Charge |
|---|---|---|---|---|
| ConfigurationBase token rates | Rate scheduleBase rate | Forecast quantity— | Billing unit— | Charge Input · Input tokens · Estimate for comparison only — / 1K tokens Output · Output tokens · Estimate for comparison only — / 1K tokens Input · Cache read · Estimate for comparison only — / 1K tokens Input · Cache write · Estimate for comparison only — / 1K tokens Input · Cache write, 1 hour · Estimate for comparison only — / 1K tokens |
| Configurationinput audio hour | Rate schedule— | Forecast quantity— | Billing unitper input audio hour | Charge Input · input audio hour · Estimate for comparison only $0.2 per input audio hour |
| ConfigurationEstimated generation cost | Rate schedule— | Forecast quantityone generation estimate | Billing unit— | Charge Output · Estimated generation cost · Estimate for comparison only $0.00005555555555555556 estimate before generation This estimate is not an additional charge or a fixed rate for the completed generation. |
Ailoot reserves an estimated amount before a request. After the result, it settles the final charge from measured usage. Streaming settlement uses the final usage record. How reservation and settlement work →
set -eu\n: "${AILOOT_API_KEY:?Set your Ailoot API key}"\n: "${AILOOT_AUDIO_FILE:?Set the path to your actual WAV recording}"\npython3 - <<'PY_REQUEST'\nimport json,os\nfrom pathlib import Path\nimport base64\naudio=Path(os.environ['AILOOT_AUDIO_FILE']).read_bytes()\nif not (audio[:4]==b'RIFF' and audio[8:12]==b'WAVE'): raise SystemExit('Supply an actual WAV audio file')\nif len(audio)>26214400: raise SystemExit('Use a documented supported large-file flow')\nrequest={'model': 'microsoft/mai-transcribe-2', 'input_audio': {'data': '${BASE64_ACTUAL_WAV_BYTES}', 'format': 'wav'}}\nrequest['input_audio']['data']=base64.b64encode(audio).decode('ascii')\nPath('request.json').write_text(json.dumps(request,ensure_ascii=False),encoding='utf-8')\nPY_REQUEST\ncurl --fail-with-body --silent --show-error https://api.ailoot.ai/v1/audio/transcriptions \\\n -H "Authorization: Bearer $AILOOT_API_KEY" -H 'Content-Type: application/json' \\\n --data-binary @request.json -D response-headers.txt -o response.bin\n# Keep the complete original response and headers. Read every returned segment, word and usage field.\n# Preserve actual operation/task/asset IDs; continue reads of the original result without resubmitting.\nGET /v1/media/tasks/{task_id}GET /v1/media/assets/{asset_id}/contentGET /v1/media/assetsGET /v1/media/assets/{asset_id}set -eu\n: "${AILOOT_API_KEY:?Set your Ailoot API key}"\ncurl --fail-with-body --silent --show-error https://api.ailoot.ai/v1/media/assets \\\n -H "Authorization: Bearer $AILOOT_API_KEY" -o owned-assets.json\n# Preserve every returned asset and follow its exact signed content URL. Do not construct a signed token.\nKeep every complete returned response, file and content part.
Retain the complete response and headers, every actual output, original returned identifiers and all usage fields. Recover the original operation without submitting creation again.
Keep the returned X-Client-Request-ID for reconciliation. Read only actual returned task_id/asset_id values or the owned media library; never build a task URL from the operation ID and never resubmit creation.
| Field | Type |
|---|---|
Fieldtext | Type |
Fieldusage | Type |
Fieldsegments[] | Type |
Fieldwords[] | Type |
Fieldspeaker fields | Type |
Fieldtimestamps | Type |
Ailoot Ultra microsoft/mai-transcribe-2
Use the method paths shown above.