Model routing

How a call is fulfilled

A live call is listen, interpret, speak. First the job type, then the language pair. NEO checks the models the operator registered and the score for that exact pair. Speech to speech is listed and is not what answers the call today.

Job type

Language pair

German → English

Example, German to English. The espeak voice on the NEO server is refused for German and does not speak this sentence. WhatsApp, email, and NEO Assistant use the same text chain, not these three speech steps.

What the operator registered

How the call is fulfilled

1. Job type
The job type decides which registered models can be used.
2. Language pair
The language pair decides which score applies.
3. What the operator registered
NEO reads only the models this operator registered for this job.
4. The score
Vendors and testers score each model for this job type and this pair. The stronger score goes first.
5. The model that answers
The first model answers.
6. The fallback
If it does not answer, NEO takes the next model in that same score for this pair and this job.

Listen

First: Whisper

Fallback: One model is registered for this step.

Interpret

First: GPT-4o mini

Fallback: One model is registered for this step.

Speak

First: GPT-4o mini TTS

Fallback: OpenAI TTS, ElevenLabs Multilingual v2

OpenAI stays on the paid API and stays in the evaluation list. Every open-weight model below is an evaluation row. None of them has a score yet. A live call does not use these rows until a test ranks them.

Paid API

Listen

WhisperIn this example
GPT Live Transcribe
GPT Transcribe

Speak

OpenAI TTSIn this example
GPT-4o mini TTSIn this example
ElevenLabs Multilingual v2In this example

Speech to speech

GPT-Live 1
GPT-Realtime 2.1
GPT-Realtime 2.1 mini
Deepslate Opal v2.1

NEO servers

Secure speech recognition
Secure speech synthesis

Open-weight evaluation

Listen

Listen
Model	License	Commercial	Size	When to use
faster-whisper large-v3	MIT	Yes	not measured this pass	Broad listener, including German, English, Arabic, and Turkish. Quality is not equal in every language.
Voxtral Mini 4B Realtime	Apache 2.0	Yes	17.73 GB	Pairs that include Arabic, or a language Parakeet does not cover. Turkish and Kurdish are not on its card.
Qwen3-ASR 1.7B	Apache 2.0	Yes	4.70 GB	Second listener. The publisher lists 30 languages and 22 Chinese dialects.
Parakeet TDT 0.6B	CC BY 4.0	Yes, with attribution	5.73 GB	European pairs, if a test ranks it above Voxtral. Arabic, Turkish, and Kurdish are not included.

faster-whisper large-v3
License
MIT
Commercial
Yes
Size
not measured this pass
When to use
Broad listener, including German, English, Arabic, and Turkish. Quality is not equal in every language.
Voxtral Mini 4B Realtime
License
Apache 2.0
Commercial
Yes
Size
17.73 GB
When to use
Pairs that include Arabic, or a language Parakeet does not cover. Turkish and Kurdish are not on its card.
Qwen3-ASR 1.7B
License
Apache 2.0
Commercial
Yes
Size
4.70 GB
When to use
Second listener. The publisher lists 30 languages and 22 Chinese dialects.
Parakeet TDT 0.6B
License
CC BY 4.0
Commercial
Yes, with attribution
Size
5.73 GB
When to use
European pairs, if a test ranks it above Voxtral. Arabic, Turkish, and Kurdish are not included.

Speak

Speak
Model	License	Commercial	Size	When to use
Chatterbox Multilingual	MIT	Yes	13.87 GB	Main voice, including Arabic, German, English, and Turkish. Each clip carries a Resemble watermark.
Chatterbox Mandarin	MIT	Yes	included above	Chinese, instead of the general Chatterbox model.
Chatterbox Hindi	MIT	Yes	included above	Hindi, instead of the general Chatterbox model.
VoxCPM2	Apache 2.0	Yes	not measured this pass	The publisher lists 30 languages, including Arabic, German, English, and Turkish. About 8 GB of GPU memory when loaded.
Qwen3-TTS	Apache 2.0	Yes	5.20 GB with its tokenizer	Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian. Arabic and Turkish are out.
Kokoro 82M	Apache 2.0	Yes	0.36 GB	English, Spanish, French, Hindi, Italian, Japanese, Brazilian Portuguese, Mandarin. German, Arabic, and Turkish are out.
MeloTTS	MIT	Yes	about 1.3 GB	English, Spanish, French, Chinese, Japanese, Korean. German and Arabic packs were not in the measured set.

Chatterbox Multilingual
License
MIT
Commercial
Yes
Size
13.87 GB
When to use
Main voice, including Arabic, German, English, and Turkish. Each clip carries a Resemble watermark.
Chatterbox Mandarin
License
MIT
Commercial
Yes
Size
included above
When to use
Chinese, instead of the general Chatterbox model.
Chatterbox Hindi
License
MIT
Commercial
Yes
Size
included above
When to use
Hindi, instead of the general Chatterbox model.
VoxCPM2
License
Apache 2.0
Commercial
Yes
Size
not measured this pass
When to use
The publisher lists 30 languages, including Arabic, German, English, and Turkish. About 8 GB of GPU memory when loaded.
Qwen3-TTS
License
Apache 2.0
Commercial
Yes
Size
5.20 GB with its tokenizer
When to use
Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian. Arabic and Turkish are out.
Kokoro 82M
License
Apache 2.0
Commercial
Yes
Size
0.36 GB
When to use
English, Spanish, French, Hindi, Italian, Japanese, Brazilian Portuguese, Mandarin. German, Arabic, and Turkish are out.
MeloTTS
License
MIT
Commercial
Yes
Size
about 1.3 GB
When to use
English, Spanish, French, Chinese, Japanese, Korean. German and Arabic packs were not in the measured set.

Speech to speech

Speech to speech
Model	License	Commercial	Size	When to use
Qwen2.5-Omni 7B	Apache 2.0	Yes	22.38 GB	Speech in, speech out. Voices are Chelsie and Ethan. Trial for English, Chinese, or German. Not for Arabic until a NEO test says so.
Step-Audio 2 mini	Apache 2.0	Yes	17.88 GB	English and Chinese conversation. Run this checkpoint first.
Step-Audio 2 mini Base	Apache 2.0	Yes	17.88 GB	Same English and Chinese limit. Score it beside mini.
Step-Audio 2 mini Think	Apache 2.0	Yes	17.88 GB	Same English and Chinese limit. Score it beside mini.
MiniCPM-o 2.6	Not stored in the NEO registry	Not stored in the NEO registry	Not measured this pass	In the product registry as ready when wired. A live call does not call it.
Moshi	Not stored in the NEO registry	Not stored in the NEO registry	Not measured this pass	In the product registry as ready when wired. A live call does not call it.
Kimi-Audio	Not stored in the NEO registry	Not stored in the NEO registry	Not measured this pass	In the product registry as ready when wired. A live call does not call it.
PersonaPlex 7B	Not stored in the NEO registry	Not stored in the NEO registry	Not measured this pass	In the product registry as ready when wired. A live call does not call it.
LFM2.5-Audio 1.5B	Not stored in the NEO registry	Not stored in the NEO registry	Not measured this pass	In the product registry as ready when wired. A live call does not call it.

Qwen2.5-Omni 7B
License
Apache 2.0
Commercial
Yes
Size
22.38 GB
When to use
Speech in, speech out. Voices are Chelsie and Ethan. Trial for English, Chinese, or German. Not for Arabic until a NEO test says so.
Step-Audio 2 mini
License
Apache 2.0
Commercial
Yes
Size
17.88 GB
When to use
English and Chinese conversation. Run this checkpoint first.
Step-Audio 2 mini Base
License
Apache 2.0
Commercial
Yes
Size
17.88 GB
When to use
Same English and Chinese limit. Score it beside mini.
Step-Audio 2 mini Think
License
Apache 2.0
Commercial
Yes
Size
17.88 GB
When to use
Same English and Chinese limit. Score it beside mini.
MiniCPM-o 2.6
License
Not stored in the NEO registry
Commercial
Not stored in the NEO registry
Size
Not measured this pass
When to use
In the product registry as ready when wired. A live call does not call it.
Moshi
License
Not stored in the NEO registry
Commercial
Not stored in the NEO registry
Size
Not measured this pass
When to use
In the product registry as ready when wired. A live call does not call it.
Kimi-Audio
License
Not stored in the NEO registry
Commercial
Not stored in the NEO registry
Size
Not measured this pass
When to use
In the product registry as ready when wired. A live call does not call it.
PersonaPlex 7B
License
Not stored in the NEO registry
Commercial
Not stored in the NEO registry
Size
Not measured this pass
When to use
In the product registry as ready when wired. A live call does not call it.
LFM2.5-Audio 1.5B
License
Not stored in the NEO registry
Commercial
Not stored in the NEO registry
Size
Not measured this pass
When to use
In the product registry as ready when wired. A live call does not call it.

Interpret

Usual providers

GPT-4o miniIn this example
GPT-4o
GPT-4
GPT-4 Turbo
GPT-5.6 Luna
GPT-5.6 Terra
GPT-5.6 Sol
GPT-6 Astra

NEO servers

NEO Llama 3.1
NEO Qwen3 32B
Qwen3 8B
Qwen3 4B
Llama 3.1 8B