#Model Selection Guide

Faced with an ever‑growing list of models, you don't need to memorise every name. This guide teaches you how to read the key metrics of each model, then match them to your own needs – language, real‑time requirement, hardware, timestamps … filter step by step, and only a handful of models will remain.

📌 All models are in ONNX format. When you use them for the first time in manyspeech, they are downloaded automatically.

#I. ASR (Speech Recognition) Models

#1. Six Key Metrics: Understand the Model, Then Write the Command

Every model table contains these columns. Once you understand them, you'll know what to choose and how to write the command.

Why can you ignore punctuation? (for Chinese/English scenarios)

• Because regardless of whether the model natively outputs punctuation, manyspeech by default calls a punctuation restoration model (--punc) to add punctuation to the output. You don't need to worry about Punctuation = No.

Microphone → online, file → offline
This is the recommended configuration in most cases. You can also mix them:

# Use an online model on a file – benefits: low resource usage, what you see is what you get.
# Drawback: accuracy may be slightly lower than using a dedicated offline model under the same conditions.
manyspeech asr -t online -m chunk -i file --files file1.wav file2.wav file3.wav

# Use an offline model on a microphone – benefit: better accuracy.
# Drawback: you must wait until the end of an utterance to see the result, and resource usage is slightly higher than with an online model.
manyspeech asr -t offline -m chunk -i mic

A very basic command example (assuming you have already chosen a model called some-model):

# Recognise a file (note that `-i file` cannot be omitted)
manyspeech asr -t offline -i file --files "meeting.wav" --model some-model

# Real‑time microphone recognition
manyspeech asr -t online -i mic --model some-model

⚠️ The asr subcommand must include -i, either -i file or -i mic. It cannot be omitted. If you omit --model, the program uses a built‑in default model, which may not be suitable for your scenario – we recommend always specifying a model.

#2. More Information Is Encoded in the Model Name

The model name itself is a “mini spec sheet”. Besides the metrics in the table, the name reveals additional details.

When you see distil-whisper-xxx, it is faster and smaller than a whisper-xxx of the same size – ideal for resource‑constrained scenarios.
When you see xxx-cantonese-onnx or xxx-wenetspeech-yue, it has been fine‑tuned for Cantonese and will be more accurate than a general model.
When you see xxx-onnx-opt, it usually performs better than xxx-onnx.

#3. Four‑Step Selection: Filter According to Your Needs

#Step 1: What language(s) do you mainly speak?

Look at the Languages column in the tables and filter for models that include your required language(s).

• Mandarin: Prefer models tagged with zh or Chinese
• Code‑switching Chinese/English: Look for zh-en or Chinese/English
• Cantonese: Look for yue or Cantonese (or names containing cantonese / yue)
• English: Look for en or English, or distil-whisper-*-en etc.
• Japanese/Korean/Thai/Russian etc.: Find models with the corresponding language tags (e.g. ja, ko, th, ru)
• Many languages worldwide: Look for multilingual, multi, or models supporting many languages (e.g. Whisper series supports 99‑106 languages)

A model that is specifically designed for a given language will usually achieve higher accuracy on that language. Multilingual models are convenient but may be slightly less accurate than dedicated ones.
If you see finetune or a specific suffix like -belle or -wenetspeech, it means the model has been fine‑tuned for a vertical domain (medical, conversation, dialect). If your scenario matches, give it priority.

#Step 2: Real‑time or offline?

Check the Type column:

#Step 3: What is your hardware level?

Look at the Precision suffix and the size encoded in the model name:

int8 quantised versions reduce size by 50‑75%, increase speed by 2‑4x, and typically lose less than 1% accuracy – strongly recommended.
distil-* models are faster and smaller than the original of the same size – also good for resource‑constrained scenarios.

#Step 4: Do you need to generate subtitles?

Without using a VAD model:

• Need SRT/VTT subtitles → You must select a model with Timestamps = Yes (the column marked Yes, or a name containing timestamp)
• No subtitles needed → Ignore the timestamp metric; any model works

When using a VAD model:

• The model's timestamp metric is not required. You can even use only an online model and still generate subtitles with timestamps.

You don't need to worry about punctuation – the program adds it automatically. So the Punctuation column can be completely ignored.

Extra requirement: Hotword customisation
If you want to improve recognition of specific terms (brand names, person names, technical terms), look for models with seaco in their name (SeACo‑Paraformer). They support hotword boosting.

#4. Common Command Templates (just fill in the model you selected)

# Offline file recognition
manyspeech asr -t offline -i file --files "your_audio.wav" --model your_chosen_model

# Real‑time microphone
manyspeech asr -t online -i mic --model your_chosen_model

# 2‑pass mode
manyspeech asr -t 2pass -i mic \
  --model streaming_model_name \
  --model2 offline_model_name

# Output subtitles (only if the model supports timestamps)
manyspeech asr -t offline -i file --files "audio.wav" \
  --format srt --model model_name

# Force use of int8 quantisation (if the model has an int8 version)
manyspeech asr -t offline -i file --files "audio.wav" \
  --accuracy int8 --model model_name

# Low‑end device: limit threads + lightweight model
manyspeech --threads 1 asr -t offline -i file --files "audio.wav" \
  --model lightweight_model_name

# Switch VAD (use silero in noisy environments)
manyspeech --vad silero-vad-v6-onnx asr -t online -i mic --model your_chosen_model

#II. VAD (Voice Activity Detection) Models

#1. Available Models

#2. Selection Advice

• Quiet environment: Use the default alifsmnvad-onnx.
• Noisy environment (fan, traffic, multiple people chatting): Switch to silero-vad-v6-onnx.

#3. Common Commands

# Specify a VAD model during ASR
manyspeech --vad silero-vad-v6-onnx asr -t online -i mic --model your_model

# Use the vad subcommand alone to detect speech segments
manyspeech vad -t offline -i file --files "recording.wav" --vad silero-vad-v6-onnx

#III. Punctuation Restoration Models

#1. Available Model

#2. Notes

• By default, ASR automatically enables this model to add punctuation to recognition results. Usually no manual intervention is needed.
• If you want to call it manually or test it, you can use the punc subcommand.

#3. Common Commands

# Manually restore punctuation for a text without punctuation
manyspeech punc --text "today is a nice day lets go for a picnic"

# Disable punctuation restoration in ASR (to test the model's native output)
manyspeech --punc "" asr -t offline -i file --files "audio.wav" --model model_name

#IV. AudioSep Audio Separation Models (planned)

This feature is under development. No models are available yet. The following is a preview of selection dimensions.

#4.1 Future Available Models (example)

• Vocals separation models (e.g. ONNX versions of Demucs, Spleeter)
• Accompaniment / instrument separation models

#4.2 Selection Dimensions (planned)

#4.3 Placeholder Command (future implementation)

# Example (future)
manyspeech audiosep -i file --files "mix.wav" --target vocals --output vocals.wav

#V. Frequently Asked Questions

Q: In the ASR table, Punctuation says “No”. Will the output have punctuation?
A: Yes. Because the program by default calls punctuation restoration (--punc). You don't need to worry about whether the model itself outputs punctuation.

Q: Can KV acceleration and int8 be used together?
A: Yes. Models whose names contain both int8 and kv/selfcrosskv support that combination.

Q: What is the difference between distil‑whisper and regular whisper?
A: distil‑whisper is a distilled version – smaller, faster, slightly lower accuracy but usually sufficient. It is suitable for resource‑constrained scenarios.

Q: How can I tell whether a model supports my language?
A: Look at the Languages column in the table, or infer from language codes in the model name (zh / yue / en / ja / ko etc.).

Q: What if downloading is too slow?
A: Download manually from ModelScope and place the files in your --base directory (by default, models/ under the program directory).

#VI. Summary: Selecting a Model Means Selecting Metrics

Follow this order to filter models in the list, put the resulting model name after --model, and run the command. If the result is not satisfactory, adjust the filters and try another batch of models.

#Next Steps

• Model Library - Detailed model documentation