Node.js 如何接收SIP音频并将wav流发送到节点中的Google语音识别API？

到目前为止，我一直在尝试，但它有一些令人生畏的限制（）。有什么办法吗？可能有一个类似星号的节点包装器

更详细地说，基本思想是

• 运行虚拟sip客户端，该客户端可以接收sip连接
• 将该连接的音频转换为常规wav格式
• 将wav音频流传输到Google语音API
• 有其他方法通过节点对sip流进行操作，如播放声音

---

最简单的方法是使用星号EAGI界面，将声音从stdin/stream读取到谷歌

---

然而目前谷歌语音识别api并不稳定。有时候它会停止工作，然后第二天开始工作。

这篇文章已经很老了，谷歌方面的情况似乎有了很大的改进，无论是语音处理器本身（越来越精确）还是Node.js方面，随着与谷歌云接口的语音API的定期更新

根据@arheops的建议，人们可能想看看Asterisk的EAGI和Node.js，以便获得谷歌转录的音频样本

以下EAGI bash脚本可能在这方面有所帮助（提供详细说明）：

---

希望这有帮助

sipster

是可配置的，您可以将pjsua2配置选项传递给

init（）

。这些选项可以在pjsua2文档中找到，它们没有在sipster文档中列出，因为有很多选项，这将是重复文档。假设您的“wave stream”表示“连续流”“在google文档中，您需要在googl端执行GRPC/proto缓冲区的路径。您应该查看您的api以访问音频缓冲区的字节。。。假设这些是编码的fmt&&比特率，与语音api输入兼容，您只需ArrayCopy.myAudioBytes（）&&写入您为语音打开的goog.api.channel。。。我试过了，但没用。你能分享示例代码吗？

#!/bin/bash

# Read all variables sent by Asterisk store them as an array, but won't use them
declare -a array
while read -e ARG && [ "$ARG" ] ; do
        array=(` echo $ARG | sed -e 's/://'`)
        export ${array[0]}=${array[1]}
done

# First argument is language
case "$1" in
"fr-FR" | "en-GB" | "es-ES" | "it-IT" )
  LANG=$1
  ;;
*)
  LANG=en-US
  ;;
esac

NODECMD=$(which node)

# Second argument is a timeout, in seconds. The duration to wait for voice input form the caller.
DURATION=$2
SAMPLE_RATE=8000
SAMPLE_SIZE_BYTES=2
let "SAMPLE_SIZE_BITS = SAMPLE_SIZE_BYTES * 8"

# EAGI_AUDIO_FORMAT is an asterisk variable that specifies the sample rate and
# sample size (usually 16 bits per sample) of the caller's voice stream.
# Depending on the codec used here, you can get sample rate values ranging from
# 8000Hz (e.g. G.711 uLaw) to 48000Hz (e.g. opus).
echo "GET VARIABLE EAGI_AUDIO_FORMAT"
read line
EAGI_AUDIO_FORMAT=$(echo $line | sed -r 's/.*\((.*)\).*/\1/')

# 5 seconds of audio input are gathered in ( SAMPLE_RATE / sample_size ) * 5 bytes
# - SAMPLE_RATE is set as per EAGI_AUDIO_FORMAT
# - sample_size is set to 2 (16 bits per sample)
#
# We don't do much here to adapt the sample rate, this code should be improved
case "${EAGI_AUDIO_FORMAT}" in
"slin48")
  SAMPLE_RATE=48000
  ;;
*)
  SAMPLE_RATE=8000
  ;;
esac

# Temporary file to store raw audio samples
AUDIO_FILE=/tmp/audio-${SAMPLE_SIZE_BITS}_bits-${SAMPLE_RATE}_hz-${DURATION}_sec.raw

# We use `dd` here to copy the raw audio samples we're getting from file
# descriptor 3 (this is the Enhanced version in EAGI) to the temporary file.
# The number of blocks to copy is a function of the DURATION to record audio and
# the sample rate. SAMPLE_SIZE_BYTES cannot be changed as it is assumed that each
# sample is 16 bits in size.
let "COUNT = SAMPLE_RATE * SAMPLE_SIZE_BYTES * DURATION"
# By default, dd stores blocks of 512 bytes
let "BLOCKS = COUNT / 512"
echo "exec noop \"Number of bytes to store : ${COUNT}\""
read line

echo "exec noop \"Number of dd blocks to store : ${BLOCKS}\""
read line

echo "exec playback \"beep\""
read line

dd if=/dev/fd/3 count=${BLOCKS} of=${AUDIO_FILE}
echo "exec noop \"File saved !\""

echo "exec noop \"AUDIO_FILE : ${AUDIO_FILE}\""
read line
echo "exec noop \"SAMPLE_RATE : ${SAMPLE_RATE}\""
read line
echo "exec noop \"LANG : ${LANG}\""
read line

# Submit audio to Google Cloud Speech API and get the result
export GOOGLE_APPLICATION_CREDENTIALS=/usr/local/node_programs/service_account_file.json
RES=$(${NODECMD} /usr/local/node_programs/nodejs-speech/samples/recognize.js sync ${AUDIO_FILE} -e LINEAR16 -r ${SAMPLE_RATE} -l ${LANG})

# clean up result returned from recognize.js :
# - remove new lines
# - remove 'Transcription :' header
RES=$(echo $RES | tr -d '\n' | sed -e 's/Transcription: \(.*$\)/\1/')

# Set GOOGLE_TRANSCRIPTION_RESULT variable, remove temporary file
# and continue dialplan execution
echo "set variable GOOGLE_TRANSCRIPTION_RESULT \"${RES}\""
read line

/bin/rm -f ${AUDIO_FILE}

exit 0