Speech Services

Speech terms — defined before code

Converting audio to text

Python — Speech-to-Text (from audio file)

Install: pip install azure-cognitiveservices-speech

import os
import azure.cognitiveservices.speech as speechsdk
# azure.cognitiveservices.speech is the Azure Speech SDK
# Commonly aliased as "speechsdk" for brevity


# ── Step 1: Configure the Speech service connection ────────────────────
# SpeechConfig holds your subscription key and region
# subscription=: your Azure AI Speech resource key
# region=: the Azure region e.g. "eastus", "canadacentral"
speech_config = speechsdk.SpeechConfig(
    subscription = os.environ.get("SPEECH_KEY"),
    region       = os.environ.get("SPEECH_REGION")
)

# Set the language the speaker is using (for recognition accuracy)
# Use BCP-47 format: "en-US", "fr-FR", "en-GB", "es-ES", etc.
speech_config.speech_recognition_language = "en-US"


# ── Step 2: Configure the audio source ─────────────────────────────────
# AudioConfig tells the SDK where to get the audio from
# from_wav_file(): read from a .wav file on disk
# from_default_microphone_input(): use the system's default microphone
audio_config = speechsdk.AudioConfig(
    filename = "meeting_recording.wav"
)


# ── Step 3: Create the SpeechRecognizer ────────────────────────────────
# SpeechRecognizer combines the speech config + audio config
# It is the main object that performs speech recognition
recognizer = speechsdk.SpeechRecognizer(
    speech_config = speech_config,
    audio_config  = audio_config
)


# ── Step 4: Recognize speech ────────────────────────────────────────────
# recognize_once() processes a single utterance (one sentence/phrase)
# It listens until a pause is detected, then returns the result
# For continuous long audio, use start_continuous_recognition() instead
result = recognizer.recognize_once()

# ── Step 5: Handle the result ───────────────────────────────────────────
# result.reason tells you what happened:
# ResultReason.RecognizedSpeech — audio was successfully transcribed
# ResultReason.NoMatch — audio detected but speech not recognizable
# ResultReason.Canceled — error occurred
if result.reason == speechsdk.ResultReason.RecognizedSpeech:
    print(f"Transcription: {result.text}")  # result.text is the transcribed text

elif result.reason == speechsdk.ResultReason.NoMatch:
    print("No speech recognised in the audio.")

elif result.reason == speechsdk.ResultReason.Canceled:
    # Get cancellation details to understand the error
    cancellation = speechsdk.CancellationDetails.from_result(result)
    print(f"Error: {cancellation.reason} — {cancellation.error_details}")

Converting text to natural-sounding audio

Python — Text-to-Speech with Neural Voice

import os
import azure.cognitiveservices.speech as speechsdk

speech_config = speechsdk.SpeechConfig(
    subscription = os.environ.get("SPEECH_KEY"),
    region       = os.environ.get("SPEECH_REGION")
)

# Set the voice to use for speech synthesis
# Voice name format: "Language-Region-VoiceName"
# Full list: https://learn.microsoft.com/azure/ai-services/speech-service/language-support
speech_config.speech_synthesis_voice_name = "en-US-AriaNeural"

# Configure audio OUTPUT — where to save or play the audio
# from_audio_file_output(): write to a .wav file
# For speakers: AudioOutputConfig(use_default_speaker=True)
audio_output = speechsdk.audio.AudioOutputConfig(
    filename = "output_speech.wav"
)

# SpeechSynthesizer combines config + audio output
synthesizer = speechsdk.SpeechSynthesizer(
    speech_config = speech_config,
    audio_config  = audio_output
)

# ── Option 1: Plain text synthesis ────────────────────────────────────
result = synthesizer.speak_text_async(
    "Hello! Welcome to the AI-102 study hub. Good luck on your exam!"
).get()  # .get() waits for the async operation to complete

if result.reason == speechsdk.ResultReason.SynthesizingAudioCompleted:
    print("Audio saved to output_speech.wav")


# ── Option 2: SSML for precise control ────────────────────────────────
# SSML (Speech Synthesis Markup Language) gives fine-grained control
# over how the text is spoken — rate, pitch, pauses, emphasis
ssml_text = """
<speak version="1.0" xmlns="http://www.w3.org/2001/10/synthesis" xml:lang="en-US">
    <voice name="en-US-AriaNeural">
        <prosody rate="slow" pitch="+5%">
            This announcement is very important.
        </prosody>
        <break time="500ms"/>
        Please listen carefully.
    </voice>
</speak>
"""
# speak_ssml_async() for SSML input instead of plain text
result_ssml = synthesizer.speak_ssml_async(ssml_text).get()

SSML elements — exam cheat sheet

Translating spoken audio in real time

Speech Translation converts spoken words from one language directly to text (or speech) in another language — in real time. It's different from using Translator on text — it handles the audio directly.

import os
import azure.cognitiveservices.speech as speechsdk

# SpeechTranslationConfig extends SpeechConfig with translation settings
translation_config = speechsdk.translation.SpeechTranslationConfig(
    subscription  = os.environ.get("SPEECH_KEY"),
    region        = os.environ.get("SPEECH_REGION"),
    speech_recognition_language = "en-US"   # The SPOKEN input language
)

# Add one or more target languages to translate INTO
# These are text output languages (ISO codes without region: "fr", "es", "de")
translation_config.add_target_language("fr")   # French
translation_config.add_target_language("es")   # Spanish

# Optional: set a voice for speech OUTPUT (translated audio)
# If set, the service also synthesises the translated text into audio
translation_config.voice_name = "fr-FR-DeniseNeural"

audio_config = speechsdk.AudioConfig(filename="english_speech.wav")

# TranslationRecognizer performs speech recognition + translation
translator = speechsdk.translation.TranslationRecognizer(
    translation_config = translation_config,
    audio_config       = audio_config
)

result = translator.recognize_once()

if result.reason == speechsdk.ResultReason.TranslatedSpeech:
    print(f"Original:  {result.text}")
    # result.translations is a dictionary: {"fr": "Bonjour...", "es": "Hola..."}
    for lang, translation in result.translations.items():
        print(f"  {lang.upper()}: {translation}")

Identifying who is speaking

Improving recognition for your domain

Standard speech-to-text models are trained on general language. If your application involves domain-specific terminology (medical, legal, technical), unusual proper nouns, strong accents, or noisy environments — a Custom Speech model significantly improves accuracy.

Custom Speech vs Standard Speech — when to use which

Choosing the right STT service

Both services transcribe speech to text — the exam tests when to use which.

Speech Services — what the exam tests

Test your understanding