How to Automatically Detect Language in Speech Using Python

Learn how to detect the language spoken in any audio, whether a recorded file or a live microphone stream, in about 20 lines of Python, using Bat Spoken Language Identification. Bat Spoken Language Identification processes audio entirely on-device, so voice data never leaves the machine and language detection works offline. The same code runs on Linux, macOS, Windows, and Raspberry Pi, and separate SDKs cover Android, iOS, and Web.

Install the Language Identification Python SDK

With Python 3.9 or newer, install pvbat and pvrecorder Python packages using PIP:

pip3 install pvbat pvrecorder

pvbat contains the spoken language identification engine and a built-in model for multiple languages. pvrecorder captures the microphone audio for identifying the spoken language in real time.

Get Your AccessKey

Sign up for a Picovoice Console account. Once your account is ready, copy your AccessKey from the Console home page.

Detect the Language of an Audio File

Start with the imports, and a small map of language codes to names:

import argparse
import struct
import wave

import pvbat

LANGUAGE_NAMES = {
    "de": "German", "en": "English", "es": "Spanish", "fr": "French",
    "it": "Italian", "ja": "Japanese", "ko": "Korean", "pt": "Portuguese",
}

Take the AccessKey and the audio file as command-line arguments:

parser = argparse.ArgumentParser()
parser.add_argument("--access_key", required=True)
parser.add_argument("--wav_path", required=True)
args = parser.parse_args()

Create the engine with your AccessKey, and read in the audio. Bat analyzes audio in fixed-size frames of bat.frame_length samples at 16 kHz (bat.sample_rate), 16-bit, single-channel; Python's built-in wave and struct modules read the WAV file into raw 16-bit samples:

bat = pvbat.create(access_key=args.access_key)

with wave.open(args.wav_path, "rb") as f:
    audio = struct.unpack("%dh" % f.getnframes(), f.readframes(f.getnframes()))

Feed the audio to the engine frame by frame. Each call to .process() returns a dictionary that maps every supported language to a score between 0 and 1, or None when the frame contains no usable voice. The language with the highest score wins:

for i in range(len(audio) // bat.frame_length):
    frame = audio[i * bat.frame_length:(i + 1) * bat.frame_length]
    scores = bat.process(frame)
    if scores:
        best = max(scores, key=scores.get)
        name = LANGUAGE_NAMES.get(str(best), "Unknown")
        print(f"Detected {name} ({scores[best]:.0%} confidence)")

bat.delete()

Full Python Code to Detect the Language of an Audio File

Save the complete script as spoken-language-identification.py:

import argparse
import struct
import wave

import pvbat

LANGUAGE_NAMES = {
    "de": "German", "en": "English", "es": "Spanish", "fr": "French",
    "it": "Italian", "ja": "Japanese", "ko": "Korean", "pt": "Portuguese",
}

parser = argparse.ArgumentParser()
parser.add_argument("--access_key", required=True)
parser.add_argument("--wav_path", required=True)
args = parser.parse_args()

bat = pvbat.create(access_key=args.access_key)

with wave.open(args.wav_path, "rb") as f:
    audio = struct.unpack("%dh" % f.getnframes(), f.readframes(f.getnframes()))

for i in range(len(audio) // bat.frame_length):
    frame = audio[i * bat.frame_length:(i + 1) * bat.frame_length]
    scores = bat.process(frame)
    if scores:
        best = max(scores, key=scores.get)
        name = LANGUAGE_NAMES.get(str(best), "Unknown")
        print(f"Detected {name} ({scores[best]:.0%} confidence)")

bat.delete()

Run Language Detection Script

Run it with your AccessKey and a WAV file:

python3 spoken-language-identification.py --access_key ${ACCESS_KEY} --wav_path sample.wav

Grab test recordings in every supported language from the Bat GitHub repository, or use any WAV of your own.

Identify the Spoken Language from the Microphone in Real Time

For live audio, PvRecorder streams small chunks from the microphone. Collect them in a buffer until a full engine frame of bat.frame_length samples accumulates, process it, then slide the buffer forward by 0.75 seconds so the scores refresh continuously as someone speaks.

Full Python Code for Real-Time Spoken Language Identification

The microphone version, spoken-language-identification-mic.py, takes only the --access_key argument:

import argparse

import pvbat
from pvrecorder import PvRecorder

parser = argparse.ArgumentParser()
parser.add_argument("--access_key", required=True)
args = parser.parse_args()

bat = pvbat.create(access_key=args.access_key)

recorder = PvRecorder(frame_length=512)
recorder.start()
print("Listening... (press Ctrl+C to stop)")

buffer = []
try:
    while True:
        buffer.extend(recorder.read())
        if len(buffer) >= bat.frame_length:
            scores = bat.process(buffer[:bat.frame_length])
            buffer = buffer[int(0.75 * bat.sample_rate):]
            if scores:
                best = max(scores, key=scores.get)
                print(f"{str(best).upper()}: {scores[best]:.2f}")
            else:
                print("(no voice detected)")
except KeyboardInterrupt:
    pass
finally:
    recorder.stop()
    bat.delete()

Run Real-Time Language Identification Script

Run the python code for spoken language identification, and speak in any of the supported languages to watch the detection update live:

python3 spoken-language-identification-mic.py --access_key ${ACCESS_KEY}

PvRecorder reads the system default microphone; to pick a specific one, list devices with PvRecorder.get_available_devices() and pass its index as device_index.

How to Read the Spoken Language Identification Output

Three details make the output easier to interpret:

• Higher scores mean higher confidence. Each supported language gets a score between 0 and 1, and the top-scoring language is the prediction. A top score near 1 signals a confident match; a low top score, or scores spread across several languages, means the frame carries weaker evidence.
• None flags silence. The engine skips frames that lack detectable voice. The voice_threshold argument of pvbat.create() (default 0.4) controls how strict that filter is: raise it to reject background noise, lower it to catch quiet speech.
• UNKNOWN catches everything else. Bat scores an open set of languages, so speech outside the supported set lands on UNKNOWN instead of forcing a wrong match.

Both programs stick to the pvbat.create() defaults; the pvbat Python API docs cover every argument, including device for choosing the inference device.

Supported Languages and Accuracy

Bat Spoken Language Identification recognizes English, French, German, Italian, Japanese, Korean, Portuguese, and Spanish. Bat identifies the spoken language with 93% accuracy, about 2x fewer errors than SpeechBrain, while using 62x less memory (5 MB versus 333 MB) and 9x less CPU.

What's Next?

Detecting the language is usually the first step in a longer pipeline. The speech-to-speech translation recipe builds one: Bat Spoken Language Identification detects the spoken language, then on-device streaming speech-to-text, translation, and streaming text-to-speech turn it into speech in the target language.

Frequently Asked Questions