Universal-3 Pro (Async)

Universal-3 Pro is our most powerful Voice AI model, designed to capture the “hard stuff” that traditional ASR models struggle with. It delivers state-of-the-art accuracy for entities, rare words, and domain-specific terminology out of the box, with code switching and optional prompting for more control. It’s also our fastest model, so you get the best accuracy without sacrificing speed.

Quickstart

Get started with Universal-3 Pro using the code below. This example transcribes a pre-recorded audio file and prints the transcript text to your terminal.

$
pip install requests

1
import requests
2
import time
3
4
base_url = "https://api.assemblyai.com"
5
headers = {"authorization": "<YOUR_API_KEY>"}
6
7
data = {
8
    "audio_url": "https://assembly.ai/sports_injuries.mp3",
9
    "language_detection": True,
10
    "speech_models": ["universal-3-pro", "universal-2"]
11
}
12
13
response = requests.post(base_url + "/v2/transcript", headers=headers, json=data)
14
15
if response.status_code != 200:
16
    print(f"Error: {response.status_code}, Response: {response.text}")
17
    response.raise_for_status()
18
19
transcript_response = response.json()
20
transcript_id = transcript_response["id"]
21
polling_endpoint = f"{base_url}/v2/transcript/{transcript_id}"
22
23
while True:
24
    transcript = requests.get(polling_endpoint, headers=headers).json()
25
    if transcript["status"] == "completed":
26
        print(transcript["text"])
27
        break
28
    elif transcript["status"] == "error":
29
        raise RuntimeError(f"Transcription failed: {transcript['error']}")
30
    else:
31
        time.sleep(3)

Key capabilities

The model out of the box outperforms all ASR models on the market on accuracy, especially as it pertains to entities and rare words. With prompting, you can get an entirely customized transcription output that rivals near-human-level transcription.

• Keyterm Prompting: Improve recognition of domain-specific terminology, rare words, and proper nouns
• Prompting: Guide transcription style, formatting, and output characteristics

To fine-tune to your use case, see the Prompting section. Not sure where to start? Use one of the recommended prompts and tweak from there.

Start with no prompt

We strongly recommend testing with no prompt first. When you omit the prompt parameter, Universal-3 Pro automatically applies a built-in default prompt that is already optimized for accuracy across a wide range of audio types — including verbatim transcription, multilingual code-switching, and challenging audio conditions. For most use cases, the default prompt delivers excellent results out of the box.

If you’re going to build a prompt, start with one of the recommended prompts and then tweak it for your use case. You should not start from scratch with your prompt — use a recommended prompt and then build off of it. Please read the Prompting Guide (Async) if you’d like to build your prompt yourself.

Remember, prompts are primarily instructional, so adding a large amount of context may not make a significant impact on accuracy and could reduce instruction-following coherence. Feel free to layer in additional instructions that you see in the Prompting Guide (Async).

Benchmarking

For more details on evaluating transcription accuracy, including tips on using semantic WER and handling substitution artifacts, see Evaluating your prompts in the Prompting Guide.

Keyterms prompting

Keyterms prompting allows you to provide up to 1,000 words or phrases (maximum 6 words per phrase) using the keyterms_prompt parameter to improve transcription accuracy for those terms and related variations or contextually similar phrases.

Here is an example showing how you can use keyterms prompting to improve transcription accuracy for a name with distinctive spelling and formatting.

Without keyterms prompting:

Hi, this is Kelly Byrne Donahue

With keyterms prompting:

Hi, this is Kelly Byrne-Donoghue

1
import requests
2
import time
3
4
base_url = "https://api.assemblyai.com"
5
headers = {"authorization": "<YOUR_API_KEY>"}
6
7
data = {
8
    "audio_url": "https://assemblyaiassets.com/audios/keyterms_prompting.wav",
9
    "language_detection": True,
10
    "speech_models": ["universal-3-pro", "universal-2"],
11
    "keyterms_prompt": ["Kelly Byrne-Donoghue"]
12
}
13
14
response = requests.post(base_url + "/v2/transcript", headers=headers, json=data)
15
16
if response.status_code != 200:
17
    print(f"Error: {response.status_code}, Response: {response.text}")
18
    response.raise_for_status()
19
20
transcript_response = response.json()
21
transcript_id = transcript_response["id"]
22
polling_endpoint = f"{base_url}/v2/transcript/{transcript_id}"
23
24
while True:
25
    transcript = requests.get(polling_endpoint, headers=headers).json()
26
    if transcript["status"] == "completed":
27
        print(transcript["text"])
28
        break
29
    elif transcript["status"] == "error":
30
        raise RuntimeError(f"Transcription failed: {transcript['error']}")
31
    else:
32
        time.sleep(3)

Remove audio tags

Universal-3 Pro generates rich transcripts that can include inline annotations such as audio event markers (e.g., [laughter], [music]) and speaker cues. If your workflow requires clean, undecorated text, set remove_audio_tags to "all" to strip all inline annotations from the transcript output.

This is especially useful for:

• Pipelines that parse transcript text downstream (NLP, search indexing, LLM input)
• Display contexts where annotations would confuse end users
• Any workflow that expects plain text output

1
import requests
2
import time
3
4
base_url = "https://api.assemblyai.com"
5
headers = {"authorization": "<YOUR_API_KEY>"}
6
7
data = {
8
    "audio_url": "https://assembly.ai/sports_injuries.mp3",
9
    "language_detection": True,
10
    "speech_models": ["universal-3-pro", "universal-2"],
11
    "remove_audio_tags": "all"
12
}
13
14
response = requests.post(base_url + "/v2/transcript", headers=headers, json=data)
15
16
if response.status_code != 200:
17
    print(f"Error: {response.status_code}, Response: {response.text}")
18
    response.raise_for_status()
19
20
transcript_response = response.json()
21
transcript_id = transcript_response["id"]
22
polling_endpoint = f"{base_url}/v2/transcript/{transcript_id}"
23
24
while True:
25
    transcript = requests.get(polling_endpoint, headers=headers).json()
26
    if transcript["status"] == "completed":
27
        print(transcript["text"])
28
        break
29
    elif transcript["status"] == "error":
30
        raise RuntimeError(f"Transcription failed: {transcript['error']}")
31
    else:
32
        time.sleep(3)

Prompting

For a comprehensive guide on crafting effective prompts, including best practices, prompt capabilities, and example prompts, see the Prompting guide.

Universal-3 Pro delivers great accuracy out of the box. To fine-tune transcription results to your use case, provide a prompt with up to 1,500 words of context in plain language. This helps the model consistently recognize domain-specific terminology, apply your preferred formatting conventions, handle code switching between languages, and better interpret ambiguous speech.

The following is our recommended prompt for verbatim multi-lingual transcription:

1
import requests
2
import time
3
4
base_url = "https://api.assemblyai.com"
5
headers = {"authorization": "<YOUR_API_KEY>"}
6
7
data = {
8
    "audio_url": "https://assembly.ai/sports_injuries.mp3",
9
    "language_detection": True,
10
    "speech_models": ["universal-3-pro", "universal-2"],
11
    "prompt": "Required: Preserve the original language(s) and script as spoken, "
12
              "including code-switching and mixed-language phrases.\n\n"
13
              "Mandatory: Preserve linguistic speech patterns including disfluencies, "
14
              "filler words, hesitations, repetitions, stutters, false starts, "
15
              "and colloquialisms in the spoken language.\n\n"
16
              "Always: Transcribe speech with your best guess based on context in "
17
              "all possible scenarios where speech is present in the audio."
18
}
19
20
response = requests.post(base_url + "/v2/transcript", headers=headers, json=data)
21
22
if response.status_code != 200:
23
    print(f"Error: {response.status_code}, Response: {response.text}")
24
    response.raise_for_status()
25
26
transcript_response = response.json()
27
transcript_id = transcript_response["id"]
28
polling_endpoint = f"{base_url}/v2/transcript/{transcript_id}"
29
30
while True:
31
    transcript = requests.get(polling_endpoint, headers=headers).json()
32
    if transcript["status"] == "completed":
33
        print(transcript["text"])
34
        break
35
    elif transcript["status"] == "error":
36
        raise RuntimeError(f"Transcription failed: {transcript['error']}")
37
    else:
38
        time.sleep(3)

Default prompt

When no prompt is provided, Universal-3 Pro automatically applies the following default prompt:

Transcribe with context and proper nouns preserved, where speech is
present in the audio. Each language as spoken. English as English.
Non-native speakers.

You can override the default prompt by providing your own prompt value. See the Prompting guide for detailed examples covering verbatim transcription, output formatting, entity accuracy, code switching, and more.

Always: Transcribe code-switching speech with your best guess based on
context in all possible scenarios where speech is present in the audio.
Languages: English, Spanish, German, French, Portuguese, Italian.
Language codes: en, es, de, fr, pt, it.

Always: Transcribe speech with your best guess based on context in all
possible scenarios where speech is present in the audio.

Required: Preserve the original language(s) and script as spoken,
including code-switching and mixed-language phrases.
Mandatory: Preserve linguistic speech patterns including disfluencies,
filler words, hesitations, repetitions, stutters, false starts, and
colloquialisms in the spoken language.
Always: Transcribe speech with your best guess based on context in all
possible scenarios where speech is present in the audio.

Transcribe this audio

Best practices for prompt engineering

See the Prompting guide for recommended prompts, capability “knobs” you can turn, and guidance on evaluating prompts against your own audio.

Support for 99 languages

With the speech_models parameter, you can list multiple speech models in priority order, allowing our system to automatically route your audio based on language support.

Model routing behavior: The system attempts to use the models in priority order falling back to the next model when needed. For example, with ["universal-3-pro", "universal-2"], the system will try to use universal-3-pro for languages it supports (English, Spanish, Portuguese, French, German, and Italian), and automatically fall back to Universal-2 for all other languages. This ensures you get the best performing transcription where available while maintaining the widest language coverage.